Evaluation method and system for coordinated development and protection of watersheds based on deep learning

Abstract

This application relates to the technical field of coordinated watershed evaluation and discloses a deep learning-based evaluation method and system for coordinated development and protection of watersheds. The method includes: collecting and preprocessing multidimensional watershed data; enhancing the dataset; inputting the dataset and samples into a ResNet-BILSTM model to extract features; constructing a GR-BILSTM integrated model to calculate evaluation indicators; performing disturbance analysis and scenario simulation to derive sensitivity results and trends; and formulating a coordinated watershed management plan. Based on a precise policy framework based on disturbance analysis and scenario simulation, this application achieves dynamic evaluation and prediction of coordinated development and protection of watersheds.

Description

Technical Field

[0001] The present application relates to the technical field of collaborative evaluation of watersheds, and in particular to an evaluation method and system for collaborative development and protection of watersheds based on deep learning. Background Art

[0002] As complex natural and social ecosystems, the coordinated development and protection and management of river basins have always been key research areas in the water resources field. Traditional river basin evaluation methods rely primarily on indicator system construction and multi-criteria decision-making analysis, such as the Analytic Hierarchy Process (AHP) and the Fuzzy Comprehensive Evaluation Method (Fuzzy Comprehensive Evaluation Method). With the advancement of computer technology, mathematical models have been widely used in river basin system analysis, including hydrological models, water quality models, and ecological models. In recent years, machine learning methods have gradually been introduced into the field of river basin management. For example, support vector machines and random forests have been used for river basin water quality prediction and land use change analysis. However, these methods often focus on a single factor or simple linear relationships and fail to fully capture the complexity of river basin systems. At the same time, deep learning technology has made breakthroughs in environmental science, especially convolutional neural networks (CNNs) and long short-term memory networks (LSTMs), which have demonstrated powerful capabilities in processing spatiotemporal data.

[0003] However, existing technologies still have obvious shortcomings in the evaluation of coordinated development and protection of river basins. Traditional evaluation methods have difficulty dealing with nonlinear and complex relationships in river basin systems, especially finding it difficult to accurately capture the nonlinear response mechanism between pollution levels and emission sources. Secondly, existing machine learning models perform poorly when processing high-dimensional heterogeneous data, and it is difficult to simultaneously integrate multi-source data such as hydrology, water quality, ecology, and socio-economics. Thirdly, data scarcity seriously restricts model performance, especially data for key scenarios such as extreme hydrological events and water pollution events are generally insufficient. Fourthly, existing evaluation methods lack systematic sensitivity analysis and scenario simulation capabilities, making it difficult to provide scientific decision-making support for coordinated management of river basins. Fifthly, most methods focus on static evaluation, lack a grasp of the dynamic evolution of river basin systems, and cannot achieve forward-looking management. These technical deficiencies have seriously limited the accuracy and practicality of the evaluation of coordinated development and protection of river basins. Summary of the Invention

[0004] This application provides an evaluation method and system for coordinated development and protection of watersheds based on deep learning, which is used for a precise policy-making framework based on disturbance analysis and scenario simulation, and realizes dynamic evaluation and prediction of coordinated development and protection of watersheds.

[0005] In the first aspect, the present application provides an evaluation method for coordinated development and protection of watersheds based on deep learning, and the evaluation method for coordinated development and protection of watersheds based on deep learning includes: collecting multi-dimensional data of the watershed and preprocessing it to obtain a multi-dimensional data set of the watershed; performing data enhancement on the multi-dimensional data set of the watershed to obtain a watershed system data sample; inputting the multi-dimensional data set of the watershed and the watershed system data sample into a ResNet-BILSTM fusion model to obtain watershed system characteristics; constructing a GR-BILSTM integration model based on the watershed system characteristics to obtain evaluation index values ​​for coordinated development and protection of the watershed; performing disturbance analysis and scenario simulation on the evaluation index values ​​to obtain watershed system sensitivity assessment results and development trends; formulating watershed coordinated policy recommendations based on the watershed system sensitivity assessment results and development trends to obtain a watershed coordinated management plan.

[0006] In a second aspect, the present application provides an evaluation system for coordinated development and protection of watersheds based on deep learning, the evaluation system for coordinated development and protection of watersheds based on deep learning comprising:

[0007] The processing module is used to collect and preprocess the multi-dimensional data of the watershed to obtain a multi-dimensional data set of the watershed;

[0008] An enhancement module, configured to perform data enhancement on the watershed multi-dimensional dataset to obtain a watershed system data sample;

[0009] An input module is used to input the watershed multi-dimensional dataset and the watershed system data sample into the ResNet-BILSTM fusion model to obtain watershed system features;

[0010] A construction module is used to construct a GR-BILSTM integrated model based on the watershed system characteristics to obtain evaluation index values ​​for coordinated development and protection of the watershed;

[0011] A simulation module is used to perform disturbance analysis and scenario simulation on the evaluation index values ​​to obtain the sensitivity assessment results and development trends of the watershed system;

[0012] The evaluation module is used to formulate basin coordinated policy recommendations based on the sensitivity assessment results and development trends of the basin system and obtain a basin coordinated management plan.

[0013] In a third aspect, a deep learning-based evaluation device for coordinated development and protection of watersheds 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 to enable the deep learning-based evaluation device for coordinated development and protection of watersheds to execute the above-mentioned deep learning-based evaluation method for coordinated development and protection of watersheds.

[0014] In a fourth aspect, a computer-readable storage medium is provided, in which instructions are stored. When the computer-readable storage medium is run on a computer, the computer executes the above-mentioned deep learning-based evaluation method for coordinated development and protection of watersheds.

[0015] In the technical solution provided by this application, a multi-dimensional data set of the basin is obtained by collecting and preprocessing the multi-dimensional data of the basin, and a watershed system data sample is obtained by using data enhancement technology. The two are input into the ResNet-BILSTM fusion model to extract the characteristics of the basin system. Based on this, the GR-BILSTM integrated model is constructed to calculate the evaluation index value, and the evaluation index value is subjected to perturbation analysis and scenario simulation. Finally, the basin collaborative policy recommendations are formulated, which has significant technical effects. This method effectively solves the problem of basin data scarcity through the preprocessing and data enhancement of multi-dimensional data sets, especially the expansion of key scene data such as extreme hydrological events and water pollution events, making model training more comprehensive and reliable, and improving the comprehensiveness and accuracy of the evaluation. Secondly, the ResNet-BILSTM fusion model is adopted in combination with the advantages of deep learning. The residual connection of the ResNet structure effectively solves the gradient vanishing problem in the deep network and improves the spatial feature extraction capability. The BILSTM structure can simultaneously consider the temporal information of the past and the future, enhances the ability to capture long-term dependencies, and makes the extraction of basin system features more accurate. Third, the GR-BILSTM integrated model seamlessly integrates data augmentation and feature extraction by fusing a generative adversarial network with a ResNet-BILSTM model. The two-stage training strategy ensures effective optimization of model parameters and improves the reliability of evaluation metric calculations. Fourth, the perturbation analysis and scenario simulation components provide a deep understanding of the response mechanisms of the watershed system. By systematically adjusting key parameters and analyzing response changes, sensitive factors and key influencing mechanisms are identified, providing a scientific basis for precise policy implementation. Fifth, the coordinated watershed management plan formulated based on the assessment results includes phased development goals, key regulatory measures, water resource management strategies, spatial control plans, and ecological restoration plans, making it highly targeted and operational. GAN data augmentation addresses data scarcity, the ResNet-BILSTM fusion structure addresses spatiotemporal feature extraction, and the GR-BILSTM integrated model addresses model integration and optimization. This allows for the accurate capture and expression of the nonlinear relationships between hydrological, water quality, and ecological factors within complex watershed systems, enabling a scientific evaluation of coordinated development and protection of the watershed. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] 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.

[0017] Figure 1 This is a schematic diagram of an embodiment of a method for evaluating coordinated development and protection of a watershed based on deep learning in an embodiment of the present application;

[0018] Figure 2 This is a schematic diagram of an embodiment of an evaluation system for coordinated development and protection of a watershed based on deep learning in an embodiment of the present application;

[0019] Figure 3 It is a schematic block diagram of the structure of an evaluation device for coordinated development and protection of a watershed based on deep learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The embodiments of the present application provide an evaluation method and system for coordinated development and protection of watersheds based on deep learning. 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 interchangeable 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.

[0021] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiments of the present application, an embodiment of the evaluation method for coordinated development and protection of watersheds based on deep learning includes:

[0022] Step S101: Collect and pre-process multi-dimensional data of the watershed to obtain a multi-dimensional data set of the watershed;

[0023] Step S102: Perform data enhancement on the watershed multi-dimensional dataset to obtain watershed system data samples; Step S103: Input the watershed multi-dimensional dataset and the watershed system data samples into the ResNet-BILSTM fusion model to obtain watershed system features;

[0024] Step S104: constructing a GR-BILSTM integrated model based on the watershed system characteristics to obtain evaluation index values ​​for coordinated development and protection of the watershed;

[0025] Step S105: Perform disturbance analysis and scenario simulation on the evaluation index values ​​to obtain the watershed system sensitivity assessment results and development trends;

[0026] Step S106: Formulate basin collaborative policy recommendations based on the basin system sensitivity assessment results and development trends to obtain a basin collaborative management plan.

[0027] It is understandable that the execution subject of this application can be an evaluation system for coordinated development and protection of watersheds based on deep learning, 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.

[0028] Specifically, multidimensional watershed data is collected and preprocessed to produce a multidimensional watershed dataset. This step specifically involves obtaining watershed hydrological data, water quality monitoring data, water body parameter data, meteorological data, water environment carrying capacity data, and ecosystem function data. Hydrological data includes runoff, water level, precipitation, and evaporation data from each monitoring station within the watershed; water quality monitoring data includes indicators such as pH, dissolved oxygen, ammonia nitrogen, total phosphorus, total nitrogen, and chemical oxygen demand; water body parameter data includes water temperature, turbidity, and conductivity; meteorological data includes temperature, humidity, wind speed, and wind direction data; water environment carrying capacity data reflects the pollutant carrying capacity of water bodies; and ecosystem function data includes indicators such as biodiversity index and vegetation cover. After obtaining the raw watershed data, data cleaning is performed to identify outliers and missing values. Outliers are identified using boxplots, and data points outside the upper and lower interquartile range (1.5 times the upper and lower interquartile range) are marked as outliers. Missing values ​​are filled using multiple interpolation or time series interpolation. The cleaned data is then normalized to convert data of varying dimensions to a uniform interval, eliminating dimensionality effects. Spatiotemporal alignment is then performed to unify data of varying spatiotemporal scales to the same spatiotemporal resolution. Feature extraction is performed on the spatiotemporally consistent data to identify key features. Finally, the time series data is decomposed into trend, cyclic, and random terms, resulting in a multidimensional watershed dataset.

[0029] Data augmentation was performed on a multi-dimensional watershed dataset to obtain watershed system data samples. Specifically, a generator network consisting of four convolutional layers and two fully connected layers and a discriminator network consisting of five convolutional neural networks were constructed to form a generative adversarial network architecture. The generator network generates simulated data, with each convolutional layer followed by a batch normalization layer and a LeakyReLU activation function. The discriminator network distinguishes between real and generated data, with each layer followed by a LeakyReLU activation function. The multi-dimensional watershed dataset was fed into the discriminator network for training to maximize the accuracy of distinguishing between real and generated samples, thereby obtaining the discriminator's optimized parameters. Based on the discriminator's optimized parameters, random noise data was fed into the generator network for training to minimize the probability of the generated samples being identified as fake, thereby obtaining the generator's optimized parameters. Using these generator and discriminator optimized parameters, the Wasserstein distance was calculated and a gradient penalty term was introduced to address training instability. Time labels and spatial location information were fed as conditional variables into the stably trained generative adversarial network, resulting in a conditional generative adversarial network. Supplementary data for extreme hydrological events and water pollution events are generated through conditional generative adversarial networks, the original data set is expanded, and watershed system data samples are obtained.

[0030] The multi-dimensional watershed dataset and watershed system data samples were input into a ResNet-BILSTM fusion model to obtain watershed system features. The ResNet-BILSTM fusion model consists of a residual network and a bidirectional long short-term memory network. A ResNet architecture consisting of five residual blocks was constructed. Each residual block contains two convolutional layers, a batch normalization layer, and a ReLU activation function, resulting in a ResNet module. The skip connection structure of the residual blocks effectively addresses the vanishing gradient problem in deep networks. A BILSTM module was constructed, consisting of a three-layer bidirectional LSTM structure, with 128 hidden units per layer and a dropout rate of 0.3. The BILSTM module was then connected to the ResNet module via a feature transfer interface, creating a hybrid architecture with skip connections, resulting in the ResNet-BILSTM fusion model. The multi-dimensional watershed dataset and watershed system data samples were combined and input into the ResNet component of the model. Feature extraction was performed through 7×7 convolutional layers and max pooling layers to obtain spatial feature vectors. The spatial feature vector is input into the BILSTM module and processed by LSTM units in both the forward and backward directions to obtain a temporal feature representation. Based on the fully connected output layer of the ResNet-BILSTM fusion model, the temporal feature representation is mapped to the feature space. Optimization training is performed using a weighted combination of mean squared error and mean absolute percentage error to obtain the watershed system characteristics.

[0031] A GR-BILSTM ensemble model was constructed based on watershed system characteristics to obtain evaluation indicators for coordinated development and protection of the watershed. The generative adversarial network data augmentation module was connected to the ResNet-BILSTM fusion model via a data stream interface to construct a GAN-ResNet-BILSTM architecture, resulting in the GR-BILSTM ensemble model. The watershed system characteristics were divided into training and validation sets, with the data split in an 8:2 ratio to obtain model training and validation data. The Adam optimizer with an initial learning rate of 0.001 was used to update the parameters of the GR-BILSTM ensemble model. The learning rate was reduced by 0.9 times every 50 training cycles to obtain a training optimization strategy. The training optimization strategy was applied to the GR-BILSTM ensemble model using the training data. An early stopping mechanism was initiated when the loss on the validation set did not improve for 10 consecutive cycles, resulting in the completed GR-BILSTM ensemble model. The trained GR-BILSTM ensemble model was used to calculate evaluation indicators such as the water quality integrity index, ecosystem service function value, water resource utilization efficiency, and system synergy, and obtain evaluation scores for each dimension. Assign corresponding weights to the evaluation scores of each dimension and perform weighted summation, normalize the results to the score range of 0-100, and obtain the evaluation index value of coordinated development and protection of the watershed.

[0032] Perturbation analysis and scenario simulations were performed on the evaluation index values ​​to obtain the results and development trends of the watershed system sensitivity assessment. Agricultural activity intensity, urban development intensity, industrial development intensity, water resource utilization intensity, and ecological protection intensity were identified as key parameters for the perturbation analysis, resulting in a perturbation analysis parameter set. Each parameter in the perturbation analysis parameter set was adjusted positively and negatively at multiple percentage levels to form a systematic parameter change scheme, resulting in a multidimensional perturbation parameter combination. The multidimensional perturbation parameter combination was applied to the evaluation index values, and multiple iterative analyses were performed to obtain data on the impact of parameter changes on the evaluation index. Based on this impact data, the ratio between parameter changes and index changes was analyzed to obtain the sensitivity index of each parameter to the system. Various watershed development scenarios were constructed, including a baseline scenario, an ecological priority scenario, an economic priority scenario, a collaborative optimization scenario, and an extreme climate scenario. Parameters were assigned to each scenario to generate scenario simulation schemes. The scenario simulation schemes were imported into a time series forecasting tool to analyze the short-, medium-, and long-term indicator trends. These were presented through spatial visualization to obtain the results and development trends of the watershed system sensitivity assessment.

[0033] Based on the results and development trends of the basin system sensitivity assessment, recommendations for coordinated basin policies are formulated, resulting in a coordinated basin management plan. Based on the results and development trends of the basin system sensitivity assessment, a system of short-term, medium-term, and long-term goals for coordinated basin development and protection is established, resulting in phased development goals. Based on the phased development goals, differentiated regulatory measures are formulated for key factors with high sensitivity, resulting in a key regulatory plan. Based on water resource-related indicators from the basin system sensitivity assessment results, a water resource optimization allocation plan based on ecological flow assurance is formulated, resulting in a water resource management strategy. Based on the spatial distribution of development trends, the basin is divided into functional zones, and land use control standards are established for different functional zones, resulting in a spatial control plan. Based on the ecological and environmental indicators from the basin system sensitivity assessment results, ecological corridor construction, wetland restoration, and soil and water conservation projects are planned and implemented, resulting in an ecological restoration plan. By integrating the key regulatory plan, water resource management strategy, spatial control plan, and ecological restoration plan, a cross-regional coordination mechanism and multi-sectoral collaboration framework are established, resulting in a coordinated basin management plan.

[0034] In an embodiment of the present application, a multi-dimensional data set of the watershed is obtained by collecting and preprocessing the multi-dimensional data of the watershed, and a watershed system data sample is obtained by using data enhancement technology. The two are input into the ResNet-BILSTM fusion model to extract the watershed system characteristics. Based on this, a GR-BILSTM integrated model is constructed to calculate the evaluation index value, and the evaluation index value is subjected to disturbance analysis and scenario simulation. Finally, a watershed collaborative policy recommendation is formulated, which has significant technical effects. This method effectively solves the problem of watershed data scarcity through the preprocessing and data enhancement of multi-dimensional data sets, especially the expansion of key scene data such as extreme hydrological events and water pollution events, making model training more comprehensive and reliable, and improving the comprehensiveness and accuracy of the evaluation. Secondly, the ResNet-BILSTM fusion model is adopted in combination with the advantages of deep learning. The residual connection of the ResNet structure effectively solves the gradient vanishing problem in the deep network and improves the spatial feature extraction capability, while the BILSTM structure can simultaneously consider the temporal information of the past and the future, enhances the ability to capture long-term dependencies, and makes the extraction of watershed system features more accurate. Third, the GR-BILSTM integrated model seamlessly integrates data augmentation and feature extraction by fusing a generative adversarial network with a ResNet-BILSTM model. The two-stage training strategy ensures effective optimization of model parameters and improves the reliability of evaluation metric calculations. Fourth, the perturbation analysis and scenario simulation components provide a deep understanding of the response mechanisms of the watershed system. By systematically adjusting key parameters and analyzing response changes, sensitive factors and key influencing mechanisms are identified, providing a scientific basis for precise policy implementation. Fifth, the coordinated watershed management plan formulated based on the assessment results includes phased development goals, key regulatory measures, water resource management strategies, spatial control plans, and ecological restoration plans, making it highly targeted and operational. GAN data augmentation addresses data scarcity, the ResNet-BILSTM fusion structure addresses spatiotemporal feature extraction, and the GR-BILSTM integrated model addresses model integration and optimization. This allows for the accurate capture and expression of the nonlinear relationships between hydrological, water quality, and ecological factors within complex watershed systems, enabling a scientific evaluation of coordinated development and protection of the watershed.

[0035] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0036] Obtain basin hydrological data, water quality monitoring data, water body parameter data, meteorological data, water environment carrying capacity data, and ecosystem function data to obtain original basin data;

[0037] Perform data cleaning on the original watershed data, identify and process outliers and missing values, and obtain cleaned watershed data;

[0038] Performing normalization processing on the cleaned watershed data to obtain normalized watershed data;

[0039] Perform spatiotemporal alignment on the normalized watershed data to obtain spatiotemporally consistent watershed data;

[0040] Perform feature extraction on the temporally and spatially consistent watershed data to obtain watershed data features;

[0041] The time series data in the basin data characteristics are decomposed, and the hydrological time series data are decomposed into trend items, period items and random items to obtain a multi-dimensional data set of the basin.

[0042] Specifically, basin hydrological data, water quality monitoring data, water parameter data, meteorological data, water environment carrying capacity data, and ecosystem function data are obtained to generate raw basin data. Basin hydrological data refers to a data set reflecting the state of the basin's hydrological cycle, including water quantity indicators such as flow, water level, precipitation, and evaporation, obtained through regular observations at hydrological monitoring stations. Water quality monitoring data refers to indicators reflecting water quality, including physical and chemical indicators such as pH, dissolved oxygen, ammonia nitrogen, total phosphorus, total nitrogen, and chemical oxygen demand, obtained through automatic water quality monitoring stations or manual sampling and analysis. Water parameter data includes physical properties such as water temperature, turbidity, and conductivity. Meteorological data includes climatic conditions such as temperature, precipitation, humidity, and wind speed within the basin, obtained through meteorological station monitoring. Water environment carrying capacity data refers to indicators of the basin's water's ability to carry pollutants, calculated through water environment capacity. Ecosystem function data includes biodiversity indices, vegetation cover, and aquatic life survey results, obtained through field surveys and remote sensing data analysis. After obtaining raw data, data cleaning is required to identify and address outliers and missing values, resulting in cleaned watershed data. Data cleaning involves removing noise, redundancy, and inconsistencies from the data through a series of processing techniques to improve data quality. In watershed data processing, statistical methods are used to identify outliers. For example, boxplots are used to calculate the upper quartile (Q3) and lower quartile (Q1), determine the interquartile range (IQR) = Q3 - Q1, and mark data points outside the range [Q1 - 1.5 IQR, Q3 + 1.5 IQR] as outliers. Outlier handling can be achieved through deletion, replacement, or smoothing. Deletion directly removes outlier data points; replacement replaces outliers with the mean, median, or a specific percentile; and smoothing uses techniques such as moving averages or local regression to smooth out outliers. Missing values ​​are filled using interpolation techniques such as linear interpolation, multiple interpolation, or time series interpolation. Linear interpolation linearly calculates the value of missing points by using data from adjacent time points; multiple interpolation estimates missing values ​​by creating multiple complete data sets and analyzing their respective results; time series interpolation considers the trend and seasonal characteristics of the time series for interpolation.

[0043] Normalization is performed on the cleaned watershed data to obtain normalized watershed data. Normalization is the process of converting data of different dimensions and orders of magnitude into a unified interval, eliminating the impact of dimension and order of magnitude differences on subsequent analysis. Common normalization methods include minimum-maximum normalization, Z-score normalization, and decimal scaling normalization. Minimum-maximum normalization linearly transforms the data to the interval [0,1]. Decimal scaling normalization scales the data to the interval [-1,1] by moving the decimal point. In watershed data processing, appropriate normalization methods can be selected for different types of data, such as hydrological and water quality data. For example, for flow data with large variations, logarithmic transformation is used followed by Z-score normalization. The normalized watershed data is then spatiotemporally aligned to obtain spatiotemporally consistent watershed data. Spatiotemporal alignment is a key step in resolving the problem of mismatch between multi-source watershed data in the temporal and spatial dimensions. In the temporal dimension, different monitoring indicators often have different sampling frequencies. For example, hydrological data may be sampled at the hourly level, water quality data at the daily or weekly level, and meteorological data at the minute level. Time alignment unifies these data to the same time scale, achieved through upsampling or downsampling. Upsampling increases the frequency of data points, such as converting daily data to hourly data. Interpolation methods are typically used; downsampling reduces the frequency of data points, such as converting hourly data to daily data. Aggregation methods such as averaging, summing, or maximum are typically used. In the spatial dimension, the spatial distribution of different monitoring sites is uneven, necessitating the use of spatial interpolation methods to generate uniform spatial grid data. Common spatial interpolation methods include inverse distance weighting (IDW), kriging, and spline interpolation. IDW uses the inverse distance from the target point as a weight to calculate interpolation; kriging considers the spatial correlation of regional variables for optimal interpolation; and spline interpolation interpolates by fitting a smooth surface.

[0044] Feature extraction is performed on temporally and spatially consistent watershed data to obtain watershed data features. Feature extraction extracts information representing the essential characteristics of the data from the raw data, reducing the data dimension while retaining key information. Principal component analysis (PCA) and autoencoders are primarily used for feature extraction in watershed data processing. PCA is a linear dimensionality reduction method that transforms potentially correlated variables into linearly uncorrelated principal components through an orthogonal transformation. The first few principal components that explain the majority of the variance are selected as extracted features. The specific steps include calculating the data covariance matrix, decomposing the covariance matrix to obtain eigenvalues ​​and eigenvectors, selecting the eigenvectors corresponding to the first k largest eigenvalues ​​to form a projection matrix, and projecting the original data into a new space. Autoencoders are a nonlinear dimensionality reduction method based on neural networks. They learn a compact representation of the data through an encoder-decoder structure. The encoder compresses the input data into a low-dimensional latent space, while the decoder attempts to reconstruct the original data from this low-dimensional representation. The training objective is to minimize reconstruction error. Autoencoders can capture the nonlinear structure of data and are particularly suitable for feature extraction in complex watershed systems. Decomposing time series data within watershed data features, hydrological time series data is decomposed into trend, cyclic, and random terms, resulting in a multidimensional watershed dataset. Time series decomposition is the process of breaking down time series data into components with specific physical meanings. The trend term reflects the long-term trend of the data, the cyclic term reflects the cyclical pattern of data changes, and the random term represents random fluctuations after removing the trend and cyclical components. Common time series decomposition methods include classical decomposition, X-12-ARIMA decomposition, and STL decomposition. The classical decomposition method assumes that the time series is an additive or multiplicative combination of trend, seasonal, and random components; the X-12-ARIMA decomposition method is a seasonal adjustment method that combines the ARIMA model with the moving average method; and the STL decomposition method is a seasonal decomposition method based on locally weighted regression, characterized by strong robustness and wide applicability. For watershed hydrological time series, such as daily average flow data, STL decomposition can extract long-term trend changes, intra-annual seasonal variations, and short-term random fluctuations, helping to understand the changing patterns and influencing factors of watershed hydrological processes.

[0045] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0046] Construct a generator network consisting of 4 convolutional layers and 2 fully connected layers and a discriminator network consisting of 5 convolutional neural networks to obtain a generative adversarial network architecture;

[0047] The multi-dimensional data set of the watershed is input into the discriminator network for training to maximize the accuracy of distinguishing between real samples and generated samples and obtain the discriminator optimization parameters;

[0048] Based on the discriminator optimization parameters, random noise data is input into the generator network for training to minimize the probability that the generated samples are identified as false, and the generator optimization parameters are obtained;

[0049] By optimizing the parameters of the generator and the discriminator, the Wasserstein distance is calculated and the gradient penalty term is introduced to obtain a stable generative adversarial network.

[0050] The time label and spatial position information are input as conditional variables into the stable training generative adversarial network to obtain the conditional generative adversarial network;

[0051] Supplementary data for extreme hydrological events and water pollution events are generated through conditional generative adversarial networks to obtain watershed system data samples.

[0052] Specifically, a generator network consisting of four convolutional layers and two fully connected layers and a discriminator network consisting of five convolutional neural networks were constructed, resulting in a generative adversarial network architecture. The generator network is a model that maps random noise to synthetic data. Its structure consists of an input layer, intermediate processing layers, and an output layer. The input layer receives a low-dimensional random noise vector, typically a 100-dimensional Gaussian distributed random number. The intermediate processing layer includes four convolutional layers, each responsible for extracting and generating higher-dimensional features from low-dimensional features. The first convolutional layer uses 64 5×5 convolutional kernels with a stride of 2 to expand the input features to a higher dimension. The second convolutional layer uses 128 5×5 convolutional kernels with a stride of 2. The third convolutional layer uses 256 5×5 convolutional kernels with a stride of 2. The fourth convolutional layer uses 512 5×5 convolutional kernels with a stride of 2. Each convolutional layer is followed by a batch normalization layer and a LeakyReLU activation function. The batch normalization layer stabilizes the training process, while the LeakyReLU activation function allows small gradients for negative values ​​to avoid neuron death. Two fully connected layers are connected after the convolutional layers. The first fully connected layer has 1024 neurons, and the second layer has the same number of neurons as the output data dimensions. The final layer uses the Tanh activation function to map the output to the range [-1, 1], generating synthetic data with the same format as real-world watershed data.

[0053] The discriminator network is a binary classification model used to distinguish between real watershed data and data synthesized by the generator. Its architecture consists of a five-layer convolutional neural network. The first layer uses 64 5×5 convolution kernels with a stride of 2; the second layer uses 128 5×5 convolution kernels with a stride of 2; the third layer uses 256 5×5 convolution kernels with a stride of 2; and the fourth layer uses 512 5×5 convolution kernels with a stride of 2. The fifth layer is a fully connected layer that outputs a scalar value representing the probability that the input data is real. Each convolution layer is followed by a LeakyReLU activation function, and the output value is finally compressed to the range [0, 1] using a Sigmoid function, representing the discriminator's confidence in the authenticity of the input data. The watershed multidimensional dataset is fed into the discriminator network for training to maximize the accuracy of distinguishing real from generated samples. Optimizing the discriminator parameters is the first step in GAN training. The processed watershed multidimensional dataset is labeled as real samples, while samples generated from random noise are labeled as fake samples. The objective function of the discriminator is to maximize classification accuracy, that is, to maximize the probability of identifying real samples as true and the probability of identifying generated samples as false. For the input real watershed data x and the generated false data G(z), the objective function of the discriminator D is expressed as: maximizing E[log(D(x))] + E[log(1-D(G(z)))], where E represents the expectation, D(x) represents the probability of the discriminator outputting a real sample as true, and D(G(z)) represents the probability of the discriminator outputting a generated sample as true. The discriminator parameters are updated using the gradient ascent method, gradually improving its ability to distinguish between true and false samples, resulting in the optimal discriminator parameters.

[0054] Based on the discriminator's optimized parameters, random noise data is fed into the generator network for training. The generator's optimized parameters are obtained by minimizing the probability that the generated samples will be identified as fake. When the discriminator's parameters are fixed, the generator's goal is to generate samples that can deceive the discriminator, that is, to make the discriminator mistakenly classify the generated samples as real samples. The generator's objective function is expressed as: minimizing E[log(1-D(G(z)))], which is equivalent to maximizing E[log(D(G(z)))]. By updating the generator parameters through gradient descent, the generator gradually improves its ability to generate realistic watershed data, and the generator's optimized parameters are obtained. The training process adopts an alternating optimization strategy. The discriminator is trained several times with the generator parameters fixed, and then the generator is trained once with the discriminator parameters fixed. This cycle repeats until the preset number of iterations is reached or the convergence condition is met.

[0055] Standard generative adversarial network training often encounters instability issues such as mode collapse or vanishing gradients. To address these issues, we optimize the parameters of the generator and discriminator, calculate the Wasserstein distance, and introduce a gradient penalty term, resulting in a stable training generative adversarial network. The Wasserstein distance is a measure of the difference between two probability distributions and is more suitable for measuring the distributional discrepancy between generated and real data than traditional JS divergence or KL divergence. With the introduction of the Wasserstein distance, the objective function of the generative adversarial network becomes: an expression that the generator attempts to minimize and the discriminator attempts to maximize. This expression is equal to the expected output of the discriminator for real data minus the expected output of the discriminator for generated data. The discriminator no longer outputs a probability value, but instead outputs a real-number score representing the sample's authenticity. To ensure that the discriminator meets certain mathematical constraints (i.e., the 1-Lipschitz constraint), a gradient penalty term is introduced. This term is the expected value of the square of the difference between the norm of the discriminator's gradient and 1 at randomly interpolated points between the real and generated samples. The final objective function is the Wasserstein distance that the generator minimizes and the discriminator maximizes minus the weighted value of the gradient penalty term, where the weight coefficient is usually set to 10. Through this improvement, the generative adversarial network training process becomes more stable, and the generator is able to produce more diverse and realistic watershed data samples.

[0056] Time labels and spatial location information are input as conditional variables into a stably trained generative adversarial network, resulting in a conditional generative adversarial network (CGN). A CGN is an extension of a GAN that allows for additional conditional information to control the generation process. In watershed data generation, time labels can control the temporal characteristics of the generated data, such as seasonal variations; spatial location information can control the spatial distribution of the generated data, such as upstream-downstream relationships. In implementation, the conditional information is concatenated with random noise as input to the generator, and concatenated with generated or real samples as input to the discriminator. The objective function of the CGN becomes an expression that the generator attempts to minimize and the discriminator attempts to maximize. This expression consists of two parts: the expected value of the logarithmic probability that the discriminator correctly identifies a real sample under the given conditions, and the expected value of the logarithmic probability that the discriminator correctly rejects a generated sample under the given conditions. The discriminator output represents the probability that a real sample is real under the given conditions, while the generator generates samples from noise under the given conditions. The training process of a CGN is similar to that of a standard GAN, except that the conditional information is added to the input.

[0057] A conditional generative adversarial network (CGN) is used to generate supplementary data for extreme hydrological and water pollution events, yielding a watershed system data sample. Extreme hydrological events such as floods and droughts, as well as severe water pollution events such as sudden pollutant exceedances, are often rare in historical monitoring data, resulting in imbalanced training data. Using the trained CGN, synthetic data resembling these rare events is generated by setting specific conditioning variables, such as time tags for extreme weather conditions or spatial location information near pollution sources. The generation process begins by setting conditioning variables based on target event characteristics, such as time tags for floods or spatial coordinates near pollution sources. A random noise vector is then generated. The conditioning variables and the random noise are fed into the conditional generator to generate synthetic data under the corresponding conditions. The generated data is post-processed to ensure compliance with physical laws and constraints, such as the acceptable range of water quality parameters. Finally, the synthetic data is merged with the original multi-dimensional watershed dataset to yield an expanded watershed system data sample.

[0058] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0059] Construct a ResNet structure consisting of five residual blocks. Each residual block contains two convolutional layers, a batch normalization layer, and a ReLU activation function, resulting in a ResNet structure module.

[0060] Construct a BILSTM part consisting of a 3-layer bidirectional LSTM structure, with 128 hidden units and a dropout rate of 0.3 in each layer, to obtain a BILSTM structural module.

[0061] The ResNet structure module and the BILSTM structure module are connected through the feature transfer interface to build a hybrid architecture with skip connections to obtain the ResNet-BILSTM fusion model;

[0062] The multi-dimensional watershed dataset is combined with the watershed system data samples and then input into the ResNet part of the ResNet-BILSTM fusion model. Feature extraction is performed through a 7×7 convolutional layer and a maximum pooling layer to obtain a spatial feature vector.

[0063] The spatial feature vector is input into the BILSTM part of the ResNet-BILSTM fusion model and processed by LSTM units in both forward and backward directions to obtain the temporal feature representation;

[0064] Based on the fully connected output layer of the ResNet-BILSTM fusion model, the time series feature representation is mapped to the feature space, and a weighted combination of mean square error and mean absolute percentage error is used for optimization training to obtain the watershed system characteristics.

[0065] Specifically, a ResNet structure consisting of five residual blocks was constructed. Each residual block contains two convolutional layers, a batch normalization layer, and a ReLU activation function, resulting in a ResNet structural module. The residual network (ResNet) is a deep convolutional neural network with skip connections that can effectively solve the gradient vanishing problem in deep network training. In watershed data processing, the ResNet structure is mainly used to extract the spatial features of watershed data. The core of ResNet is the residual block, each of which contains two convolution operation paths and an identity mapping path. Specifically, the first path of each residual block contains two 3×3 convolutional layers, each followed by a batch normalization layer and a ReLU activation function; the second path is an identity mapping that directly passes the input to the output. The outputs of the two paths are added together and then activated by the ReLU function to obtain the final output.

[0066] In this application, the ResNet structure contains a total of 5 residual blocks, and the number of channels of each residual block is 64, 128, 256, 512 and 512 respectively. The construction process defines the residual block class, which contains two convolutional layers, a batch normalization layer and a ReLU activation function, as well as a jump connection structure. Then, the five residual blocks are stacked in sequence to form a complete ResNet structure module. The residual blocks are downsampled through a convolutional layer with a stride of 2, gradually reducing the spatial dimension of the feature map and increasing the number of channels. The input of the ResNet structure module is multi-channel watershed data, and the output is a high-dimensional feature vector that represents the spatial characteristics of the watershed data.

[0067] The BILSTM component is constructed, consisting of a three-layer bidirectional LSTM structure, with 128 hidden units per layer and a dropout rate of 0.3, resulting in a BILSTM structural module. The Long Short-Term Memory (LSTM) network is a special type of recurrent neural network designed to process long-term dependencies in sequential data. The bidirectional LSTM (BILSTM) considers both forward and backward information in the sequence, more comprehensively capturing contextual dependencies. In watershed data processing, the BILSTM is primarily used to extract dynamic features from time series data and capture temporal dependencies within watershed systems.

[0068] Each BILSTM unit consists of an input gate, a forget gate, an output gate, and a cell state. The input gate controls the extent to which new information enters the cell state; the forget gate controls the extent to which old information is retained; and the output gate controls the extent to which the cell state is transferred to the output; the cell state stores long-term memory. The bidirectional structure includes a forward LSTM and a backward LSTM. The forward LSTM processes data from the beginning of the sequence to the end, while the backward LSTM processes data from the end to the beginning. The outputs of both directions are combined to obtain the final output. The BILSTM structure in this method consists of three layers, each with 128 hidden units. A dropout rate of 0.3 is added between layers to prevent overfitting. The input of the BILSTM module is a time series feature sequence, and the output is a fixed-dimensional time series feature representation.

[0069] The ResNet and BILSTM modules are connected via a feature transfer interface to construct a hybrid architecture with skip connections, resulting in a ResNet-BILSTM fusion model. The feature transfer interface is the key structure connecting the two different types of networks, responsible for adapting the spatial features extracted by the ResNet into the temporal features required by the BILSTM. Specifically, the high-dimensional feature vector output by the ResNet is compressed in spatial dimensions through a global average pooling layer to produce a fixed-length feature vector. The feature dimensions are then adjusted through a fully connected layer to match the BILSTM input dimensions. To preserve temporal information, the feature vector is replicated multiple times to form a temporal sequence, or feature vectors from different time points are concatenated into a sequence. Finally, the processed feature sequence is input into the BILSTM for temporal feature extraction.

[0070] To enhance the model's feature representation capabilities, a skip connection structure is added to the fusion model. Features are directly extracted from the ResNet intermediate layers and, after processing, concatenated with the output of the BILSTM intermediate layers to form a multi-scale feature fusion. This hybrid architecture combines the spatial feature extraction capabilities of CNNs with the temporal modeling capabilities of RNNs, making it particularly suitable for processing complex data in watershed systems that are both spatially distributed and temporally evolving.

[0071] The multi-dimensional watershed dataset and the watershed system data samples are combined and fed into the ResNet component of the ResNet-BILSTM fusion model. Feature extraction is performed through a 7×7 convolutional layer and a maximum pooling layer to obtain a spatial feature vector. The preprocessed multi-dimensional watershed dataset is merged with the watershed system data samples generated by the GAN in the previous step to form an enhanced training dataset. During the data combination process, it is important to ensure that the formats of the two data components are consistent, including dimensionality, range, and normalization. The combined data is fed into the first layer of the ResNet-BILSTM fusion model, a 7×7 convolutional layer. This convolutional layer uses 64 7×7 convolutional kernels with a stride of 2 to extract low-level features. This convolutional layer is followed by a batch normalization layer and a ReLU activation function, followed by a 3×3 maximum pooling layer with a stride of 2 to further reduce the feature map resolution while retaining salient features. These low-level processing units capture local spatial patterns and texture information in the watershed data, such as water quality distribution gradients or hydrological connectivity patterns. The processed feature map is then passed through five residual blocks, ultimately outputting a high-dimensional feature map. The feature map is compressed into a spatial feature vector of fixed dimension through global average pooling or fully connected layer to represent the spatial feature distribution of the watershed system.

[0072] The spatial feature vector is input into the BILSTM portion of the ResNet-BILSTM fusion model and processed by LSTM units in both the forward and backward directions to obtain a time series feature representation. After being converted through the feature transfer interface, the spatial feature vector is input into the BILSTM architecture. In the BILSTM, data is processed simultaneously from two directions: the forward LSTM processes data from the start point to the end point of the time series, capturing the influence of past information on the current state; the backward LSTM processes data from the end point to the start point of the time series, capturing the influence of future information on the current state. The first layer of the BILSTM receives the transferred spatial feature sequence, calculates the hidden state through LSTM units in both directions, and concatenates the forward and backward hidden states as the output. The second and third layers of the BILSTM further extract more complex time-dependent features. To prevent overfitting, a dropout layer with a dropout rate of 0.3 is added after each BILSTM layer to randomly drop neurons to enhance the model's generalization ability. The output of the final BILSTM layer represents the time series feature representation extracted from the watershed data, containing information about the dynamic characteristics of the watershed system over time.

[0073] Based on the fully connected output layer of the ResNet-BILSTM fusion model, the time series feature representation is mapped to the feature space. A weighted combination of mean squared error and mean absolute percentage error is used for optimization training to obtain the watershed system characteristics. The time series feature representation is mapped to the final feature space through one or more fully connected layers. The number of neurons in each fully connected layer decreases, and the output dimension of the last layer matches the target feature dimension. The fully connected layer maps high-dimensional features to a specific feature space through a learned weight matrix.

[0074] During model training, a weighted combination of mean squared error (MSE) and mean absolute percentage error (MAPE) is used as the loss function. MSE measures the sum of the squared differences between the predicted and true values ​​and is sensitive to outliers; MAPE measures the percentage of relative error and is insensitive to numerical magnitude. This weighted combination balances the impact of absolute and relative errors, providing an appropriate training target for evaluation metrics of varying magnitudes. The loss function is expressed as L = α·MSE + (1-α)·MAPE, where α is a weighting factor, typically set to 0.5. The optimizer uses the Adam algorithm, with an initial learning rate of 0.001 and a learning rate decay strategy that reduces the learning rate by a factor of 0.9 every 50 epochs. An early stopping strategy is implemented during training, stopping training after 10 consecutive epochs without improvement in the validation set loss to prevent overfitting. Model parameters are optimized using backpropagation, ultimately resulting in a ResNet-BILSTM fusion model that accurately extracts watershed system characteristics.

[0075] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0076] The conditional generative adversarial network and the ResNet-BILSTM fusion model are connected through a data stream connector to form a GAN-ResNet-BILSTM framework and obtain the GR-BILSTM integrated model.

[0077] The GR-BILSTM integrated model is trained using a two-stage training method. In the first stage, the conditional generative adversarial network parameters are fixed, and only the ResNet-BILSTM fusion model is trained to obtain the first-stage training results.

[0078] Based on the results of the first phase of training, we perform end-to-end joint training, adjust the parameters of the conditional generative adversarial network and the ResNet-BILSTM fusion model, and add a regularization term for the generated data quality to obtain the trained GR-BILSTM integrated model.

[0079] Construct a multi-level evaluation framework that includes the ecological environment dimension, resource utilization dimension, and system coordination dimension, set specific evaluation indicators for each dimension, and obtain an evaluation indicator system;

[0080] The watershed system characteristics were input into the trained GR-BILSTM integrated model to calculate the water quality integrity index, biodiversity index, water resource utilization efficiency, and system coordination, obtaining evaluation results for each dimension.

[0081] The analytic hierarchy process is used to calculate the weight of each indicator, and the evaluation results of each dimension are weighted and summed and normalized to the range of 0-100 to obtain the evaluation index value of coordinated development and protection of the watershed.

[0082] Specifically, a conditional generative adversarial network (GAN) and a ResNet-BILSTM fusion model are connected via a data stream connector to form a GAN-ResNet-BILSTM framework, resulting in a GR-BILSTM integrated model. The data stream connector is a specially designed model interface layer that coordinates data transfer and feature fusion between different deep learning modules. In this method, the data stream connector consists of three main components: a feature mapping layer, a feature fusion layer, and a parameter sharing mechanism. The feature mapping layer, consisting of a series of fully connected layers, transforms the data feature space generated by the conditional generative adversarial network into a feature representation compatible with the input layer of the ResNet-BILSTM fusion model. The feature fusion layer employs an attention mechanism to dynamically weightedly fuse the generated data features with the original data features, enhancing the model's ability to learn key features. The parameter sharing mechanism establishes a partially shared parameter layer between the two models to promote consistency and complementarity in feature representation. Through the synergy of these three components, the watershed system data samples generated by the conditional generative adversarial network can be seamlessly integrated with the original watershed multi-dimensional dataset and jointly input into the ResNet-BILSTM model for feature extraction and pattern recognition, forming a GAN-ResNet-BILSTM framework, namely the GR-BILSTM integrated model.

[0083] The constructed GR-BILSTM ensemble model is trained using a two-stage approach. In the first stage, the parameters of the conditional generative adversarial network (GAN) are fixed, and only the ResNet-BILSTM fusion model is trained, resulting in the first-stage training results. During this stage, all parameters of the GAN are frozen and do not participate in gradient updates. During training, the original multi-dimensional watershed dataset and watershed system data samples generated by the GAN are fed into the ResNet-BILSTM fusion model via a data stream connector. The ResNet component extracts spatial features from the data, the BILSTM component extracts temporal features, and the fully connected layer maps the temporal features to the target feature space. The model output is compared with the true labels to calculate the loss function. Backpropagation is used to calculate gradients, and the Adam optimizer is used to update the parameters of the ResNet-BILSTM fusion model. The training goal of this stage is to enable the ResNet-BILSTM fusion model to fully utilize the augmented data generated by the GAN and improve its feature extraction capabilities.

[0084] Based on the results of the first phase of training, end-to-end joint training is performed. The parameters of the conditional generative adversarial network and the ResNet-BILSTM fusion model are adjusted simultaneously, and a generated data quality regularization term is added to obtain the trained GR-BILSTM integrated model. The second phase of training unfreezes the conditional generative adversarial network parameters and realizes the joint optimization of the two models. The generated data quality regularization term is added to the loss function, which is expressed as:

[0085] ;

[0086] in is the total loss function, is the prediction loss, which measures the difference between the model output and the true label; is the quality loss of generated data, which measures the authenticity and diversity of generated data; is a balancing coefficient that controls the relative importance of the two loss terms. The introduction of a generated data quality regularization term ensures that the data generated by the conditional generative adversarial network not only facilitates feature extraction from the ResNet-BILSTM but also meets the physical constraints and statistical characteristics of the watershed system data. During the end-to-end joint training process, the parameters of the two models are updated simultaneously, mutually reinforcing and optimizing each other. After multiple iterations of training, training terminates when performance on the validation set no longer improves, resulting in the final trained GR-BILSTM ensemble model.

[0087] A multi-level evaluation framework encompassing ecological, resource, and system coordination dimensions was constructed, with specific evaluation indicators set for each dimension to create an evaluation indicator system. The multi-level evaluation framework is a structured evaluation method that organizes various aspects of coordinated development and protection of a river basin into a clearly defined indicator system according to its inherent logic. The ecological, environmental, and resource utilization dimensions assess the health of the river basin's ecosystem, including indicators such as the water quality integrity index, biodiversity index, water ecosystem service function value, and hydrological connectivity index. The resource utilization dimension evaluates the efficiency of resource development and utilization in the river basin, including indicators such as water resource utilization efficiency, land resource utilization intensity, energy consumption elasticity coefficient, and the level of circular economy development. The system coordination dimension evaluates the degree of coordination between subsystems within the river basin, including indicators such as upstream and downstream coordination, regional development balance, multi-sector policy coordination, and ecological-economic system coupling. Each indicator has a clear calculation method and data source to ensure the objectivity and repeatability of the evaluation results.

[0088] The trained GR-BILSTM integrated model is fed with watershed system characteristics to calculate indicators such as the water quality integrity index, biodiversity index, water resource utilization efficiency, and system synergy, yielding evaluation results for each dimension. The GR-BILSTM integrated model uses deep learning and pattern recognition to automatically extract key features and map them to various evaluation indicators. During the calculation process, the model feeds the watershed system characteristics into a conditional generative adversarial network (CGN) to generate targeted augmented data. Both the original features and augmented data are then fed into a ResNet-BILSTM fusion model. The ResNet architecture extracts spatial features, while the BILSTM architecture extracts temporal features. Finally, multiple fully connected layers are used to map the extracted features to the value domains of the various evaluation indicators. The water quality integrity index reflects the physical and chemical properties and water quality of the watershed; the biodiversity index measures the species richness and uniformity of the watershed ecosystem; the water resource utilization efficiency assesses the economic and social benefits generated per unit of water resource consumption; and the system synergy quantifies the degree of interaction and coordination between subsystems within the watershed. The values ​​of each indicator calculated by the GR-BILSTM integrated model form the evaluation results for each dimension.

[0089] The analytic hierarchy process is used to calculate the weight of each indicator, and the evaluation results of each dimension are weighted and summed and normalized to the range of 0-100 to obtain the evaluation index value of the coordinated development and protection of the watershed. The analytic hierarchy process is a systematic method for determining indicator weights. By constructing a judgment matrix, calculating eigenvectors and performing consistency tests, the relative importance weights of each indicator are obtained. The specific steps include: establishing an indicator hierarchy to clarify the subordinate relationship between indicators at all levels; constructing a pairwise comparison judgment matrix, and having field experts score indicators according to their importance; calculating the maximum eigenvalue and corresponding eigenvector of the judgment matrix, and the eigenvector is the weight value after normalization; performing a consistency test to ensure the rationality of the judgment matrix; if the consistency requirements are not met, the judgment matrix needs to be reconstructed. After obtaining the weight value of each indicator, the evaluation results of each dimension are weighted and summed:

[0090] ;

[0091] in is the final evaluation score, D is the total number of dimensions, is the weight of the d-th dimension, is the number of indicators under the d-th dimension, is the weight of the kth indicator under the dth dimension, is the standardized value of the corresponding indicator. Finally, the weighted summation result is normalized to the range of 0-100 through linear transformation to obtain the evaluation index value of the coordinated development and protection of the watershed. Based on the evaluation index value, the level of coordinated development and protection of the watershed can be divided into five levels: 0-20 indicates severe imbalance, 21-40 indicates mild imbalance, 41-60 indicates basic coordination, 61-80 indicates good coordination, and 81-100 indicates high-quality coordination.

[0092] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0093] Determine the intensity of agricultural activities, urban development intensity, industrial development intensity, water resource utilization intensity, and ecological protection intensity as key parameters for disturbance analysis, and obtain a set of disturbance analysis parameters;

[0094] Each parameter in the disturbance analysis parameter set is adjusted positively and negatively at multiple percentage levels to form a systematic parameter change plan and obtain a multi-dimensional disturbance parameter combination;

[0095] Apply multi-dimensional perturbation parameter combinations to the evaluation index values, perform multiple iterative analyses, and obtain data on the impact of parameter changes on the evaluation index;

[0096] According to the impact data, the ratio relationship between the parameter change and the index change is analyzed to obtain the sensitivity index of each parameter to the system;

[0097] Construct a variety of basin development scenarios, including baseline scenario, ecological priority scenario, economic priority scenario, collaborative optimization scenario, and extreme climate scenario, configure corresponding parameters for each scenario, and obtain scenario simulation plans;

[0098] The scenario simulation scheme is imported into the time series forecasting tool to analyze the changing trends of indicators in the short, medium and long term, and presented through spatial visualization to obtain the sensitivity assessment results and development trends of the watershed system.

[0099] Specifically, agricultural activity intensity, urban development intensity, industrial development intensity, water resource utilization intensity, and ecological protection intensity were identified as key parameters for disturbance analysis, resulting in a set of disturbance analysis parameters. Agricultural activity intensity refers to the scale and density of agricultural production activities within a watershed, including indicators such as irrigation area, fertilizer application, and pesticide use. Irrigation area reflects agricultural water demand, while fertilizer and pesticide application are directly related to the intensity of non-point source pollution. Urban development intensity refers to the speed and scale of urbanization, including indicators such as urban land area, proportion of impervious area, and population density. Urban land area reflects changes in land use structure, the proportion of impervious area affects rainfall runoff characteristics, and population density is related to domestic sewage discharge. Industrial development intensity refers to the scale and intensity of industrial production activities, including indicators such as industrial land area, industrial wastewater discharge, and industrial energy consumption, which directly affect point source pollution loads and resource consumption. Water resource utilization intensity refers to the degree of water resource development and utilization, including indicators such as water withdrawal, water resource utilization efficiency, and water resource allocation ratio, reflecting the water balance of the watershed. Ecological protection intensity refers to the strength of ecosystem protection and restoration efforts, including indicators such as ecological water use, wetland protection area, and forest and grass coverage, reflecting the stability of the basin's ecosystem. By analyzing historical data and relevant literature, we identified five key parameters as the basis for disturbance analysis, forming a set of disturbance analysis parameters.

[0100] Each parameter in the perturbation analysis parameter set is adjusted positively and negatively at multiple percentage levels to form a systematic parameter change scheme, resulting in a multidimensional perturbation parameter combination. To comprehensively assess the impact of parameter changes on the watershed system, a reasonable perturbation range and gradient must be set. Typically, each parameter is adjusted with amplitudes of ±5%, ±10%, ±15%, ±20%, and ±25%, resulting in 11 perturbation levels. Positive perturbations simulate parameter growth scenarios, such as expansion of agricultural area and urban growth, while negative perturbations simulate parameter reduction scenarios, such as decreased industrial emissions and enhanced ecological protection. Combining these 11 perturbation levels for five key parameters theoretically yields 11^5 perturbation scenarios, but this would impose an excessively high computational burden. In practice, a perturbation scheme is constructed using a combination of single-factor perturbation methods and orthogonal experimental designs. Single-factor perturbation involves adjusting only one parameter at a time while keeping the others constant, resulting in a total of 5 × 11 = 55 basic perturbation scenarios. The orthogonal experimental design selects 25 representative multi-factor combination scenarios based on the L25 (5^6) orthogonal table. Ultimately, 80 perturbation parameter combinations are generated, covering both single-factor effects and multi-factor interactions. Multidimensional perturbation parameter combinations are applied to the evaluation index values, and multiple iterative analyses are performed to obtain data on the impact of parameter changes on the evaluation index. This step is the core of the perturbation analysis. By substituting the perturbation parameters into the GR-BILSTM integrated model, changes in the values ​​of the watershed coordinated development and protection evaluation index under various perturbation scenarios are predicted. The specific implementation process includes: starting from the baseline scenario, calculating the watershed evaluation index values ​​in the undisturbed state as a reference; then, for each perturbation scenario, the perturbed parameters are input into the GR-BILSTM integrated model; the model predicts the watershed evaluation index values ​​under that perturbation scenario based on the trained network weights; and calculating the absolute and relative changes in the evaluation index values ​​before and after the perturbation. This process is repeated until all perturbation scenarios have been analyzed. This process requires multiple iterations to ensure the stability and reliability of the results. Monte Carlo methods are typically used, randomly sampling each disturbance scenario multiple times (e.g., 100 times) to obtain a statistical distribution of changes in the evaluation indicators and avoid the random errors of a single simulation. Ultimately, a complete impact dataset is generated, encompassing the changes in the evaluation indicators under each disturbance scenario.

[0101] Based on the impact data, the ratio between parameter changes and indicator changes is analyzed to determine the sensitivity index of each parameter to the system. The sensitivity index is a key metric that quantifies the responsiveness of a watershed system to parameter changes. Common sensitivity indices include the elasticity coefficient, the sensitivity coefficient, and the relative sensitivity index. The elasticity coefficient is calculated as E = (ΔY / Y) / (ΔX / X), where Y represents the evaluation indicator value, X represents the disturbance parameter, and ΔY and ΔX represent the corresponding changes, respectively. The larger the absolute value of the elasticity coefficient, the more sensitive the watershed system is to that parameter. The sensitivity coefficient is calculated as S = ΔY / ΔX, measuring the absolute response of the indicator change to the parameter change. The relative sensitivity index, which takes into account the differences in the dimension and range of variation of different parameters, is calculated as RSI = (ΔY / Y) / (ΔX / Xmax - Xmin). The sensitivity index is calculated for each parameter and ranked from highest to lowest by absolute value to identify key sensitive factors. The sign of the sensitivity index is also considered: a positive value indicates that an increase in the parameter leads to an increase in the indicator, while a negative value indicates that an increase in the parameter leads to a decrease in the indicator. A comprehensive analysis of the sensitivity changes under different disturbance amplitudes reveals the nonlinear response characteristics of the watershed system.

[0102] Multiple basin development scenarios were constructed, including a baseline scenario, an ecological priority scenario, an economic priority scenario, a collaborative optimization scenario, and an extreme climate scenario. Parameters were assigned to each scenario to generate scenario simulation plans. Scenario simulations are based on the results of perturbation analysis to design development paths with practical decision-making implications. The baseline scenario maintains current development trends, with parameters continuing at historical rates and no additional regulatory measures. The ecological priority scenario emphasizes ecological protection, characterized by reducing the intensity of agricultural and industrial development, increasing investment in ecological protection, limiting urban expansion, and ensuring ecological flow. The economic priority scenario focuses on economic development, characterized by increasing the intensity of industrial and urban development, expanding agricultural production, and moderately relaxing environmental constraints. The collaborative optimization scenario seeks a balance between ecological protection and economic development, improving resource utilization efficiency through technological innovation, optimizing the industrial structure, and achieving coordinated development. The extreme climate scenario considers extreme hydrological events brought about by climate change, such as increased frequency of droughts and floods, and examines the stability and resilience of basin systems under extreme conditions. Based on the scenario settings, specific values ​​for five key parameters were assigned to each scenario to form a complete scenario simulation plan.

[0103] Scenario simulations are imported into a time series forecasting tool to analyze near-, medium-, and long-term indicator trends. These trends are presented through spatial visualization, yielding the results of the watershed system sensitivity assessment and development trends. Time series forecasting is the process of inferring the future state of a watershed system based on historical data and scenario parameters. In this method, a trained GR-BILSTM ensemble model is used as the forecasting tool. The forecast period is typically divided into the near-term (1-3 years), medium-term (3-5 years), and long-term (5-10 years). The forecasting process begins with an initial state, i.e., the current state of the watershed system. Scenario parameters are then imported, including the set values ​​of five key parameters for different time periods. The GR-BILSTM model is then run to predict the values ​​of watershed evaluation indicators at each time point. Finally, the forecast results for each time point are summarized to form a complete time series trend. The forecast results are presented in various visualization methods, including time series graphs, spatial distribution heat maps, and key node trajectory graphs. The time series graphs illustrate the temporal trends of each evaluation indicator; the spatial distribution heat maps display the spatial differences in indicators within the watershed; and the key node trajectory graphs track the path of state changes at important monitoring points or regions. By comparing the prediction results under different scenarios, the response characteristics of the basin system to various development paths are evaluated.

[0104] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0105] Based on the results of the basin system sensitivity assessment and development trends, a short-term, medium-term and long-term goal system for the coordinated development and protection of the basin was established, and phased development goals were obtained;

[0106] According to the phased development goals, differentiated regulatory measures are formulated for key influencing factors with high sensitivity to obtain key regulatory plans;

[0107] Based on the water resource related indicators in the sensitivity assessment results of the watershed system, a water resource optimization allocation plan based on ecological flow guarantee is formulated to obtain a water resource management strategy;

[0108] Based on the spatial distribution characteristics of development trends, the basin is divided into functional zones, and land use control standards are formulated for different functional zones to obtain a spatial control plan;

[0109] Based on the ecological and environmental indicators in the watershed system sensitivity assessment results, we plan and implement ecological corridor construction, wetland restoration and soil and water conservation projects to obtain ecological restoration plans;

[0110] Integrate key regulatory plans, water resources management strategies, spatial control plans and ecological restoration plans, establish a cross-regional coordination mechanism and a multi-department linkage framework, and obtain a basin collaborative management plan.

[0111] Specifically, based on the results of the basin system sensitivity assessment and development trends, a system of short-term, medium-term, and long-term goals for coordinated development and protection of the basin was established, resulting in phased development targets. This target system was established using a hierarchical approach, breaking down the overall goals into two dimensions: time and function. The time dimension was divided into three phases: short-term (1-3 years), medium-term (3-5 years), and long-term (5-10 years). The function dimension encompassed four aspects: ecological and environmental improvement, efficient resource utilization, economic development optimization, and system synergy. Target values ​​were set based on scenario simulation results predicted by the GR-BILSTM model, taking into account the indicator trajectory under the synergy optimization scenario. The specific process involved extracting time series data from the model predictions to determine expected indicator values ​​at key time points. The expected values ​​were then adjusted to ensure both challenging and achievable values ​​based on the basin's actual carrying capacity and development potential. The adjusted indicator values ​​were then categorized by time and function to form a structured target system matrix. Short-term goals focused on addressing prominent, sensitive issues, medium-term goals emphasized system balance and synergy, and long-term goals aimed at achieving the ideal state of sustainable development in the basin. This phased, multi-dimensional target system provides a clear development roadmap for coordinated river basin management.

[0112] Based on the phased development goals, differentiated regulatory measures are formulated for key influencing factors with high sensitivity, resulting in a key regulatory plan. This plan is based on the results of a sensitivity index analysis, prioritizing factors with high sensitivity. The specific development process includes: ranking factors from high to low by sensitivity index to identify the top-ranked key influencing factors; analyzing the sensitivity variations of these factors across different regions and time periods; and designing differentiated regulatory measures tailored to these characteristics. For factors with high positive sensitivity, such as ecological protection intensity, strengthening and promoting measures are designed; for factors with high negative sensitivity, such as industrial pollution intensity, restrictive and limiting measures are designed. The key regulatory plan also includes a tiered control strategy, with three levels of control based on sensitivity: mandatory control, guidance control, and supervisory control. Mandatory control applies to factors with extremely high sensitivity and significant impact, imposing the strictest regulations; guidance control applies to factors with moderate sensitivity, guiding their optimization through economic incentives and technical support; and supervisory control applies to factors with low sensitivity but requiring attention, implementing routine monitoring and periodic evaluation. Through differentiated control, limited management resources are focused on addressing key issues, improving regulatory efficiency.

[0113] Based on water resource-related indicators from the sensitivity assessment results of the watershed system, an optimal water resource allocation plan based on ecological flow assurance was developed, resulting in a water resource management strategy. This water resource management strategy takes ecological flow assurance as the core constraint and maximizes water resource utilization efficiency as the optimization goal. Water resource-related indicators, such as water resource development and utilization rate, water environmental carrying capacity, and ecological water use guarantee rate, were extracted from the sensitivity assessment results, and their sensitivity characteristics and threshold ranges were analyzed. Ecological flow requirements for key sections of the watershed were then determined, including two levels: basic ecological flow and ideal ecological flow. The basic ecological flow represents the minimum requirement for maintaining the basic ecological functions of the river, while the ideal ecological flow represents the recommended flow to support a healthy ecosystem. Next, a water resource optimization model was constructed based on the hydrological and water resource model and the GR-BILSTM model prediction results. This model uses maximizing socioeconomic benefits as the objective function, with constraints such as ecological flow assurance, total water resource control, and water quality compliance, comprehensively considering the needs of different water-using sectors and regions. By solving this optimization model, water resource allocation plans for different year levels were derived, including appropriate water allocation and regulatory measures for each region and sector. Water resources management strategies also include wet and dry season regulation mechanisms, inter-basin water transfer strategies and emergency response plans to ensure that the rational allocation of water resources in the basin can be maintained under various hydrological conditions.

[0114] Based on the spatial distribution of development trends, the basin is divided into functional zones, and land use control standards are formulated for each functional zone to develop a spatial management plan. This spatial management plan is a regionalized management strategy based on the spatial heterogeneity and functional differences of the basin. The development trends predicted by the GR-BILSTM model are spatially visualized to generate heat maps of the spatial distribution of each evaluation indicator. By overlaying and analyzing these heat maps, regional units with similar functions within the basin are identified. Functional zones are then delineated based on physical geographic features, administrative boundaries, and existing planning. Typical functional zones include: water conservation zone, ecological conservation zone, agricultural development zone, urban development zone, and industrial concentration zone. Secondly, differentiated land use control standards are formulated based on the dominant functions and sensitive characteristics of each functional zone. Through refined spatial management, the orderly development of different regions in accordance with their primary functional positioning is guided, thereby optimizing the spatial structure of the basin.

[0115] Based on the ecological and environmental indicators derived from the sensitivity assessment results of the watershed system, ecological corridor construction, wetland restoration, and soil and water conservation projects are planned and implemented to develop an ecological restoration plan. This ecological restoration plan is a systematic solution to the problem of ecosystem degradation in the watershed. Ecological and environmental indicators, such as the biodiversity index, water quality integrity index, and ecosystem service function value, are extracted from the sensitivity assessment results to identify ecologically vulnerable areas and key ecological nodes. Appropriate restoration projects are then planned for different types of ecological issues. Ecological corridor construction focuses on restoring the connectivity of fragmented ecosystems within the watershed. These corridors are divided into three categories: river corridors, mountain corridors, and green corridors. River corridors are constructed along major rivers and tributaries to ensure the continuity of aquatic habitats; mountain corridors connect mountain ecosystems within the watershed; and green corridors establish green belts within human activity areas to reduce landscape fragmentation. Wetland restoration projects target degraded and occupied wetlands, restoring their functions through measures such as improving hydrological conditions, controlling pollution, revegetating vegetation, and reestablishing habitats. Soil and water conservation projects implement measures such as slope terracing, vegetation restoration, and integrated small watershed management in areas of severe soil erosion to reduce soil erosion and non-point source pollution. Ecological restoration plans also include a restoration schedule and technical route selection, prioritizing restoration projects with significant ecological benefits and proven technologies. By integrating key regulatory plans, water resource management strategies, spatial control plans, and ecological restoration plans, a cross-regional coordination mechanism and a multi-sectoral collaboration framework will be established to develop a coordinated watershed management plan. This coordinated watershed management plan is a comprehensive management system that organically integrates the aforementioned specific plans into a coordinated and efficient management framework. This integration process analyzes the compatibility and synergy of each specific plan, identifying potential points of conflict and areas of synergy. Then, through spatiotemporal matching optimization, the plans are coordinated across time and space to avoid conflicts and promote synergy. Establishing a cross-regional coordination mechanism is key to achieving coordinated management of both upstream and downstream, and both left and right banks of the watershed.

[0116] The above describes the evaluation method of the coordinated development and protection of the watershed based on deep learning in the embodiment of the present application. The following describes the evaluation system of the coordinated development and protection of the watershed based on deep learning in the embodiment of the present application. Figure 2 In the embodiment of the present application, an evaluation system for coordinated development and protection of a watershed based on deep learning includes:

[0117] Processing module 201, used to collect and pre-process multi-dimensional data of the watershed to obtain a multi-dimensional data set of the watershed;

[0118] An enhancement module 202 is configured to perform data enhancement on the watershed multi-dimensional dataset to obtain a watershed system data sample;

[0119] An input module 203 is configured to input the watershed multi-dimensional dataset and the watershed system data sample into a ResNet-BILSTM fusion model to obtain watershed system features;

[0120] A construction module 204 is used to construct a GR-BILSTM integrated model based on the watershed system characteristics to obtain evaluation index values ​​for coordinated development and protection of the watershed;

[0121] The simulation module 205 is used to perform disturbance analysis and scenario simulation on the evaluation index value to obtain the watershed system sensitivity assessment results and development trends;

[0122] The evaluation module 206 is used to formulate basin collaborative policy recommendations based on the basin system sensitivity assessment results and development trends to obtain a basin collaborative management plan.

[0123] Through the collaborative efforts of these components, a multidimensional watershed data set is obtained by collecting and preprocessing it. Data augmentation techniques are then used to generate watershed system data samples. These samples are then fed into a ResNet-BILSTM fusion model to extract watershed system characteristics. Based on this, a GR-BILSTM integrated model is constructed to calculate evaluation index values. Perturbation analysis and scenario simulations are then performed on these evaluation index values. Ultimately, recommendations for coordinated watershed policy implementation are formulated, demonstrating significant technical effectiveness. This method effectively addresses the scarcity of watershed data through multidimensional data set preprocessing and data augmentation. In particular, the addition of data for key scenarios such as extreme hydrological events and water pollution incidents makes model training more comprehensive and reliable, improving the comprehensiveness and accuracy of evaluations. Secondly, the ResNet-BILSTM fusion model leverages the advantages of deep learning. The residual connections of the ResNet structure effectively address the vanishing gradient problem in deep networks, enhancing spatial feature extraction capabilities. The BILSTM structure simultaneously considers both past and future temporal information, enhancing the ability to capture long-term dependencies and enabling more accurate extraction of watershed system characteristics. Third, the GR-BILSTM integrated model seamlessly integrates data augmentation and feature extraction by fusing a generative adversarial network with a ResNet-BILSTM model. The two-stage training strategy ensures effective optimization of model parameters and improves the reliability of evaluation metric calculations. Fourth, the perturbation analysis and scenario simulation components provide a deep understanding of the response mechanisms of the watershed system. By systematically adjusting key parameters and analyzing response changes, sensitive factors and key influencing mechanisms are identified, providing a scientific basis for precise policy implementation. Fifth, the coordinated watershed management plan formulated based on the assessment results includes phased development goals, key regulatory measures, water resource management strategies, spatial control plans, and ecological restoration plans, making it highly targeted and operational. GAN data augmentation addresses data scarcity, the ResNet-BILSTM fusion structure addresses spatiotemporal feature extraction, and the GR-BILSTM integrated model addresses model integration and optimization. This allows for the accurate capture and expression of the nonlinear relationships between hydrological, water quality, and ecological factors within complex watershed systems, enabling a scientific evaluation of coordinated development and protection of the watershed.

[0124] above Figure 2 The evaluation system for coordinated development and protection of watersheds based on deep learning in an embodiment of the present invention is described in detail from the perspective of modular functional entities. The evaluation device for coordinated development and protection of watersheds based on deep learning in an embodiment of the present invention is described in detail from the perspective of hardware processing.

[0125] Figure 3This is a schematic diagram of the structure of a deep learning-based evaluation device for coordinated watershed development and protection, provided by an embodiment of the present invention. This deep learning-based evaluation device 300 can vary significantly depending on configuration or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors), memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage medium 330 may be either transient or persistent storage. The program stored in the storage medium 330 may include one or more modules (not shown), each of which may include a series of instructions for the deep learning-based evaluation device 300. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, allowing the deep learning-based evaluation device 300 to execute the series of instructions stored in the storage medium 330 to implement the steps of the deep learning-based evaluation method for coordinated watershed development and protection.

[0126] The deep learning-based watershed coordinated development and protection evaluation device 300 may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 3 The structure of the evaluation device for coordinated development and protection of watersheds based on deep learning shown does not constitute a limitation of the evaluation device for coordinated development and protection of watersheds based on deep learning provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0127] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions. When the instructions are executed on a computer, the computer executes the steps of the deep learning-based evaluation method for coordinated development and protection of watersheds.

[0128] 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.

[0129] If the integrated unit is implemented in the form of 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 part that contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a deep learning-based watershed collaborative development and protection evaluation device (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: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program code.

[0130] 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 deep learning-based evaluation method for coordinated development and protection of watersheds, characterized by: The method comprises: Collect and preprocess the multi-dimensional data of the watershed to obtain a multi-dimensional data set of the watershed; Performing data enhancement on the watershed multi-dimensional dataset to obtain a watershed system data sample; Inputting the watershed multi-dimensional dataset and the watershed system data sample into the ResNet-BILSTM fusion model to obtain watershed system features; Based on the characteristics of the watershed system, a GR-BILSTM integrated model is constructed to obtain the evaluation index values ​​of the coordinated development and protection of the watershed; Performing disturbance analysis and scenario simulation on the evaluation index values ​​to obtain the results of the watershed system sensitivity assessment and development trends, including: determining the intensity of agricultural activities, urban development intensity, industrial development intensity, water resource utilization intensity, and ecological protection intensity as key parameters for disturbance analysis to obtain a disturbance analysis parameter set; performing positive and negative adjustments on each parameter in the disturbance analysis parameter set according to multiple percentage levels to form a systematic parameter change scheme to obtain a multi-dimensional disturbance parameter combination; applying the multi-dimensional disturbance parameter combination to the evaluation index values, performing multiple iterative analyses, and obtaining the impact data of parameter changes on the evaluation index; analyzing the ratio relationship between the parameter change amount and the index change amount based on the impact data to obtain the sensitivity index of each parameter to the system; constructing a variety of watershed development scenarios including baseline scenarios, ecological priority scenarios, economic priority scenarios, collaborative optimization scenarios, and extreme climate scenarios, configuring corresponding parameters for each scenario, and obtaining a scenario simulation scheme; importing the scenario simulation scheme into a time series forecasting tool to analyze the short-term, medium-term, and long-term indicator change trends, and presenting them through spatial visualization to obtain the watershed system sensitivity assessment results and development trends; Based on the sensitivity assessment results and development trends of the watershed system, recommendations for coordinated watershed policy implementation are formulated to obtain a coordinated watershed management plan.

2. The deep learning-based evaluation method for coordinated development and protection of watersheds according to claim 1 is characterized in that: The multi-dimensional data of the watershed is collected and pre-processed to obtain a multi-dimensional data set of the watershed, including: Obtain basin hydrological data, water quality monitoring data, water body parameter data, meteorological data, water environment carrying capacity data, and ecosystem function data to obtain original basin data; Performing data cleaning on the original watershed data, identifying and processing outliers and missing values, and obtaining cleaned watershed data; performing normalization processing on the cleaned watershed data to obtain normalized watershed data; Performing spatiotemporal alignment on the normalized watershed data to obtain spatiotemporally consistent watershed data; Performing feature extraction on the temporally and spatially consistent watershed data to obtain watershed data features; The time series data in the watershed data features are decomposed, and the hydrological time series data are decomposed into trend items, period items and random items to obtain the watershed multi-dimensional data set.

3. The evaluation method for coordinated development and protection of watersheds based on deep learning according to claim 1 is characterized in that: The data enhancement of the watershed multi-dimensional dataset to obtain a watershed system data sample includes: Construct a generator network consisting of 4 convolutional layers and 2 fully connected layers and a discriminator network consisting of 5 convolutional neural networks to obtain a generative adversarial network architecture; Inputting the multi-dimensional data set of the watershed into the discriminator network for training, maximizing the discrimination accuracy between real samples and generated samples, and obtaining the discriminator optimization parameters; Based on the discriminator optimization parameters, random noise data is input into the generator network for training to minimize the probability that the generated samples are identified as false, thereby obtaining the generator optimization parameters; Utilizing the generator optimization parameters and the discriminator optimization parameters, the Wasserstein distance is calculated and a gradient penalty term is introduced to obtain a stably trained generative adversarial network; Inputting the time label and the spatial position information as conditional variables into the stably trained generative adversarial network to obtain a conditional generative adversarial network; Supplementary data for extreme hydrological events and water pollution events are generated through the conditional generative adversarial network to obtain the watershed system data samples.

4. The deep learning-based evaluation method for coordinated development and protection of watersheds according to claim 3 is characterized in that: The multi-dimensional data set of the watershed and the watershed system data samples are input into the ResNet-BILSTM fusion model to obtain watershed system features, including: Construct a ResNet structure consisting of five residual blocks. Each residual block contains two convolutional layers, a batch normalization layer, and a ReLU activation function, resulting in a ResNet structure module. Construct a BILSTM part consisting of a three-layer bidirectional LSTM structure, with 128 hidden units and a dropout rate of 0.3 per layer, to obtain a BILSTM structural module. The ResNet structure module and the BILSTM structure module are connected through a feature transfer interface to construct a hybrid architecture with skip connections to obtain the ResNet-BILSTM fusion model; Combining the watershed multi-dimensional dataset with the watershed system data sample and inputting the combined data into the ResNet part of the ResNet-BILSTM fusion model, performing feature extraction through a 7×7 convolutional layer and a maximum pooling layer to obtain a spatial feature vector; Input the spatial feature vector into the BILSTM part of the ResNet-BILSTM fusion model, and process it through LSTM units in both forward and backward directions to obtain a temporal feature representation; Based on the fully connected output layer of the ResNet-BILSTM fusion model, the time series feature representation is mapped to the feature space, and a weighted combination of mean square error and mean absolute percentage error is used for optimization training to obtain the watershed system characteristics.

5. The deep learning-based evaluation method for coordinated development and protection of watersheds according to claim 4 is characterized in that: The GR-BILSTM integrated model is constructed based on the watershed system characteristics to obtain evaluation index values ​​for coordinated development and protection of the watershed, including: Connecting the conditional generative adversarial network and the ResNet-BILSTM fusion model through a data stream connector to form a GAN-ResNet-BILSTM framework and obtain a GR-BILSTM integrated model; The GR-BILSTM integrated model is trained using a two-stage training method. In the first stage, the conditional generative adversarial network parameters are fixed, and only the ResNet-BILSTM fusion model part is trained to obtain the first-stage training results. Based on the training results of the first phase, end-to-end joint training is performed. The parameters of the conditional generative adversarial network and the ResNet-BILSTM fusion model are adjusted, and a regularization term for the generated data quality is added to obtain the trained GR-BILSTM integrated model. Construct a multi-level evaluation framework that includes the ecological environment dimension, resource utilization dimension, and system coordination dimension, set specific evaluation indicators for each dimension, and obtain an evaluation indicator system; Input the watershed system characteristics into the trained GR-BILSTM integrated model, calculate the water quality integrity index, biodiversity index, water resource utilization efficiency and system coordination, and obtain evaluation results of each dimension; The weight value of each indicator is calculated using the hierarchical analysis method, and the evaluation results of each dimension are weighted and summed and normalized to the range of 0-100 to obtain the evaluation index value of the coordinated development and protection of the watershed.

6. The deep learning-based evaluation method for coordinated development and protection of watersheds according to claim 1 is characterized in that: The basin collaborative policy recommendations are formulated based on the basin system sensitivity assessment results and development trends to obtain a basin collaborative management plan, including: Based on the results of the watershed system sensitivity assessment and the development trends, establish a short-term, medium-term and long-term target system for coordinated development and protection of the watershed, and obtain phased development goals; Based on the phased development goals, differentiated regulatory measures are formulated for key influencing factors with high sensitivity to obtain key regulatory plans; Based on the water resource related indicators in the sensitivity assessment results of the watershed system, a water resource optimization allocation plan based on ecological flow guarantee is formulated to obtain a water resource management strategy; Based on the spatial distribution characteristics of the development trend, the basin is divided into functional zones, and land use control standards are formulated for different functional zones to obtain a spatial control plan; Based on the ecological and environmental indicators in the watershed system sensitivity assessment results, plan and implement ecological corridor construction, wetland restoration and soil and water conservation projects to obtain an ecological restoration plan; Integrate the key regulatory plan, the water resources management strategy, the spatial control plan and the ecological restoration plan, establish a cross-regional coordination mechanism and a multi-department linkage framework, and obtain the basin collaborative management plan.

7. A deep learning-based evaluation system for coordinated development and protection of watersheds, characterized by: The method for implementing the deep learning-based evaluation system for coordinated development and protection of a watershed according to any one of claims 1 to 6 comprises: A processing module is used to collect and pre-process the multi-dimensional data of the watershed to obtain a multi-dimensional data set of the watershed; an enhancement module is used to perform data enhancement on the multi-dimensional data set of the watershed to obtain a watershed system data sample; An input module is used to input the watershed multi-dimensional dataset and the watershed system data sample into the ResNet-BILSTM fusion model to obtain watershed system characteristics; A construction module is used to construct a GR-BILSTM integrated model based on the watershed system characteristics to obtain evaluation index values ​​for coordinated development and protection of the watershed; A simulation module is used to perform disturbance analysis and scenario simulation on the evaluation index value to obtain the sensitivity assessment results and development trends of the watershed system, including: determining the intensity of agricultural activities, urban development intensity, industrial development intensity, water resource utilization intensity, and ecological protection intensity as key parameters for disturbance analysis to obtain a disturbance analysis parameter set; performing positive and negative adjustments on each parameter in the disturbance analysis parameter set according to multiple percentage levels to form a systematic parameter change scheme to obtain a multidimensional disturbance parameter combination; applying the multidimensional disturbance parameter combination to the evaluation index value, performing multiple iterative analyses, and obtaining the impact data of parameter changes on the evaluation index; analyzing the ratio relationship between the parameter change amount and the index change amount based on the impact data to obtain the sensitivity index of each parameter to the system; constructing a variety of watershed development scenarios including a baseline scenario, an ecological priority scenario, an economic priority scenario, a collaborative optimization scenario, and an extreme climate scenario, configuring corresponding parameters for each scenario, and obtaining a scenario simulation scheme; importing the scenario simulation scheme into a time series forecasting tool to analyze the short-term, medium-term, and long-term indicator change trends, and presenting them through spatial visualization to obtain the sensitivity assessment results and development trends of the watershed system; The evaluation module is used to formulate basin coordinated policy recommendations based on the sensitivity assessment results and development trends of the basin system and obtain a basin coordinated management plan.

8. A deep learning-based evaluation device for coordinated development and protection of watersheds, characterized by: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, it implements the evaluation method for coordinated development and protection of watersheds based on deep learning as described in any one of claims 1 to 6.

9. 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 executes the evaluation method for coordinated development and protection of watersheds based on deep learning as described in any one of claims 1 to 6.

Images