A small and micro wetland ecological restoration intelligent monitoring method and system

Abstract

The present application relates to the technical field of ecological environment monitoring and intelligent information processing, and discloses a kind of small and micro wetland ecological restoration intelligent monitoring method and system.The method comprises: through multi-source sensor array synchronous acquisition water, soil, gas, life multi-dimensional parameter, constructs the ecological state vector of space-time alignment;Cross-media nonlinear coupling modeling is carried out using graph attention spatio-temporal fusion network;Key water quality and biological indicators are predicted using a mixed architecture of LSTM and GRU;Combined with three-level health threshold and rule base, graded early warning and adaptive repair recommendations are generated.The system includes six modules: multi-source perception, data preprocessing, graph structure construction, feature fusion, time series prediction and intelligent decision-making.The system realizes global perception, early warning and intervention of small and micro wetland ecological state, and improves the timeliness and scientificity of repair.

Description

Technical Field

[0001] This invention belongs to the field of ecological environment monitoring and intelligent information processing technology, specifically relating to an intelligent monitoring method and system for the ecological restoration of small wetlands. Background Technology

[0002] With the deepening of ecological civilization construction, small wetlands, as key nodes in urban ecological networks, play an irreplaceable role in water conservation, water purification, and biodiversity maintenance. Their ecosystems are intricately structured and dynamically sensitive, making them highly vulnerable to human activities and climate change.

[0003] Currently, the ecological health assessment of small wetlands mainly relies on fixed-point sampling and laboratory analysis, supplemented by a limited number of water quality or meteorological sensors for auxiliary monitoring. While these methods can obtain local physicochemical parameters, they are insufficient to comprehensively depict the complex ecological processes involving the coupling of water, soil, and organisms within the wetland system, and cannot effectively reflect the comprehensive changes in key ecological states such as sediment organic matter cycling, microbial activity, and aquatic vegetation succession.

[0004] Ecological intelligent monitoring based on multi-source sensor data has become an important direction for improving the management efficiency of small wetlands. This technology aims to integrate remote sensing imagery, in-situ sensor networks, and environmental model outputs to build a continuous perception and dynamic inversion capability for wetland ecological parameters. Its core objective is to analyze the nonlinear evolution laws of ecosystems through high-dimensional heterogeneous data, achieving a leap from "point observation" to "area extrapolation," and providing real-time evidence for degradation early warning and restoration decisions.

[0005] In existing technologies, conventional statistical regression or shallow machine learning models are used to correlate monitoring data with ecological indicators. However, their ability to model the strong nonlinearity, high time-varying nature, and multi-factor interactions in wetland systems is severely insufficient, resulting in low accuracy in the inversion of key ecological parameters. Simultaneously, traditional threshold-based alarm mechanisms lack the ability to predict ecological degradation trends proactively, often triggering responses only after sudden changes in water quality or the collapse of biological communities, leading to significant delays in early warning. These shortcomings make it difficult for existing monitoring systems to support the refined restoration needs of small wetlands requiring "early identification and early intervention." Therefore, there is an urgent need for an intelligent monitoring method that integrates deep learning and multi-source sensing to achieve a breakthrough in the integrated technology of high-precision ecological parameter inversion and proactive anomaly early warning. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention is proposed. Embodiments of this invention provide an intelligent monitoring method and system for the ecological restoration of small wetlands. It synchronously collects multi-dimensional dynamic parameters of water bodies, soil, atmosphere, and biological communities using a multi-source heterogeneous ecological sensor array, constructing a spatiotemporally aligned high-dimensional ecological state vector. A spatiotemporal feature fusion network based on a graph attention mechanism is employed to perform cross-media coupling modeling of multi-source data, extracting nonlinear interaction features of the small wetland ecosystem. An ecological evolution prediction model based on a hybrid architecture of long short-term memory networks and gated cyclic units is introduced to perform multi-step advanced predictions of key water quality indicators and biodiversity indices. Combined with a preset ecological health threshold system and a dynamic risk assessment rule base, a graded early warning signal is generated and adaptive restoration strategy recommendations are triggered, achieving accurate perception, advanced early warning, and intelligent intervention of the ecological degradation process of small wetlands.

[0007] This invention provides an intelligent monitoring method for the ecological restoration of small wetlands, comprising:

[0008] By simultaneously acquiring water physicochemical parameters, soil environmental parameters, atmospheric meteorological parameters, and biological activity parameters through a multi-source sensor array deployed in different ecological media of small wetlands, a raw ecological perception data stream is formed.

[0009] The original ecological sensing data stream is processed by timestamp alignment, outlier removal and missing value imputation to generate a standardized multidimensional time-series ecological dataset.

[0010] Based on the multidimensional time-series ecological dataset, an ecological relationship graph structure is constructed with water bodies, soil, atmosphere, and organisms as nodes. The node features are parameter vectors corresponding to each medium, and the edge weights are determined by the physical and chemical coupling strength between the media.

[0011] The ecological relationship graph structure is input into the graph attention spatiotemporal fusion network, and the dynamic dependency weights of cross-media features are calculated through a multi-head attention mechanism to output the fused high-order ecological state representation vector.

[0012] The higher-order ecological state representation vector is input into a hybrid prediction model composed of a long short-term memory network and a gated recurrent unit to predict dissolved oxygen concentration, ammonia nitrogen content, total phosphorus concentration and benthic biodiversity index in multiple future time steps.

[0013] The predicted results are compared with the preset ecological health classification thresholds, and a level 1, 2 or 3 ecological early warning signal is generated according to the risk level judgment rules. At the same time, corresponding restoration measures suggestions such as vegetation replanting, aeration and oxygenation or microbial release are output.

[0014] Preferably, the original ecological sensing data stream is processed by timestamp alignment, outlier removal, and missing value imputation to generate a standardized multidimensional time-series ecological dataset, including:

[0015] Based on Coordinated Universal Time, all sensor data are bucketed together in the edge computing gateway, and the arithmetic mean of multiple sampled values ​​of the same parameter within the window is taken as the representative value of that time point.

[0016] For each parameter sequence, the mean and standard deviation are calculated. Data points that exceed the mean plus or minus three times the standard deviation are marked as outliers and removed.

[0017] For parameter sequences with a continuous missing duration of less than 15 minutes, piecewise cubic spline interpolation is used to fill in the missing values. If the duration is greater than 15 minutes, it is marked as an invalid period and a backup sensor data source is enabled to generate a standardized multidimensional time-series ecological dataset with a sampling frequency of once per minute.

[0018] Preferably, an ecological relationship graph structure with water bodies, soil, atmosphere, and organisms as nodes is constructed based on the multidimensional time-series ecological dataset, including:

[0019] The standardized multidimensional time-series ecological dataset is divided into four node sets according to media type: water node feature vector dimension is 7, including water temperature, pH value, dissolved oxygen, conductivity, turbidity, redox potential and total phosphorus concentration; soil node feature vector dimension is 3, including water content, temperature and organic matter content; atmosphere node feature vector dimension is 4, including air temperature, humidity, wind speed and light intensity; biological node feature vector dimension is 2, including acoustic activity index and thermal imaging biological density.

[0020] Based on water-soil exchange flux, air-water interface mass transfer coefficient and biological migration pathway, the connection relationships between nodes are established, including water-soil bidirectional connection, water-atmosphere unidirectional connection, soil-atmosphere bidirectional connection, water-biological unidirectional connection and soil-biological unidirectional connection.

[0021] The initial values ​​of the edge weights are set based on the historical measured coupling coefficients, and the edge weights are fine-tuned online based on the prediction residual feedback after each prediction cycle.

[0022] Preferably, the ecological relationship graph structure is input into a graph attention spatiotemporal fusion network, and the dynamic dependency weights of cross-media features are calculated through a multi-head attention mechanism. The fused high-order ecological state representation vector is output, including:

[0023] The ecological relationship graph structure is sequentially passed through two graph convolutional layers to extract local neighborhood features. The graph convolutional layers are implemented using Chebyshev polynomial approximation.

[0024] The output of the second graph convolutional layer is input into a multi-head graph attention layer containing 8 parallel attention heads. The query, key, and value projection matrix of each head has a dimension of 16. Attention scores are calculated by scaling dot products and then normalized by softmax to obtain the attention weights between nodes.

[0025] The outputs of each attention head are spliced, linearly transformed, residual connected, and layer normalized to generate a high-order ecological state representation vector.

[0026] Preferably, the higher-order ecological state representation vector is input into a hybrid prediction model composed of a cascaded long short-term memory network and a gated recurrent unit to predict dissolved oxygen concentration, ammonia nitrogen content, total phosphorus concentration, and benthic biodiversity index over multiple future time steps, including:

[0027] The high-order ecological state representation vector of 72 consecutive time steps is used as input, corresponding to the historical state of the past 72 hours.

[0028] Temporal dependency abstraction is performed sequentially through the first long short-term memory network layer and the second long short-term memory network layer;

[0029] The output of the second long short-term memory network layer is input into the gated recurrent unit layer to integrate long-term trend and short-term fluctuation characteristics.

[0030] The output layer, composed of four fully connected neurons, generates hourly predictions of dissolved oxygen, ammonia nitrogen, total phosphorus, and benthic biodiversity index. The activation function of the output layer is a linear function.

[0031] Preferably, the training of the hybrid prediction model adopts a weighted loss function, which is composed of the mean squared error and the mean absolute percentage error weighted in a ratio of 7:3, and an early stopping mechanism is introduced to prevent overfitting.

[0032] Preferably, the prediction results are compared with preset ecological health grading thresholds, and level 1, 2, or 3 ecological early warning signals are generated according to risk level determination rules, and corresponding remediation measures are simultaneously output, including:

[0033] A Level 3 warning is triggered when any two indicators are predicted to be less than the Level 1 threshold for three consecutive hours; a Level 2 warning is triggered when they are less than the Level 2 threshold; and a Level 1 warning is triggered when they are less than the Level 3 threshold.

[0034] Based on the warning level, the system matches the corresponding vegetation replanting type, aeration equipment start-up and shutdown sequence, compound microbial agent dosage, and ecological floating bed deployment density from the pre-set restoration strategy knowledge base to generate structured restoration instructions.

[0035] Preferably, the repair strategy knowledge base includes 12 categories of standard measures, each of which is associated with specific execution parameters, applicable conditions, and expected effects.

[0036] This invention also provides an intelligent monitoring system for the ecological restoration of small wetlands, comprising:

[0037] The multi-source ecological sensing module is used to simultaneously collect data on water temperature, pH value, dissolved oxygen, conductivity, turbidity, redox potential, soil moisture content, soil temperature, soil organic matter content, air temperature and humidity, wind speed, light intensity, acoustic biological activity index, and infrared thermal imaging biological distribution density through sensor arrays deployed in small wetland water bodies, bottom sediments, near-ground atmosphere, and vegetation canopy.

[0038] The data preprocessing module is used to align the raw data output by the multi-source ecological sensing module with a unified time reference, use the sliding window mid-value filtering method to remove abrupt outliers, and use cubic spline interpolation to fill in missing values ​​caused by sampling interruptions, forming a standardized multi-dimensional time series data matrix with a sampling frequency of once per minute.

[0039] The ecological graph structure construction module is used to divide the standardized multidimensional time series data matrix into four node sets according to the medium type. The feature dimension of each node is the number of parameters contained in the corresponding medium. The connection edges between nodes are established based on the water-soil exchange flux, air-water interface mass transfer coefficient and biological migration path. The initial value of the edge weight is determined by the historical measured coupling coefficient and is dynamically updated during operation.

[0040] The graph attention spatiotemporal fusion module is used to receive the dynamic graph sequence output by the ecological graph structure construction module, extract local neighborhood features through two graph convolutional layers, calculate the importance weights between global nodes through a multi-head graph attention layer, and finally aggregate to generate a high-order ecological state representation vector.

[0041] The hybrid time-series prediction module is used to input the higher-order ecological state representation vector into a cascaded structure containing two long short-term memory network layers and one gated recurrent unit layer in time steps. The long short-term memory network layer has 256 hidden units, the gated recurrent unit layer has 128 hidden units, and the output layer has 4 fully connected neurons, which correspond to the hourly prediction values ​​of dissolved oxygen, ammonia nitrogen, total phosphorus, and benthic biodiversity index for the next 24 hours, respectively.

[0042] The intelligent early warning and remediation decision-making module is used to compare the predicted values ​​output by the hybrid time-series prediction module with the three-level ecological health thresholds one by one. When the predicted value of any indicator is less than the first-level threshold for 3 consecutive hours, a third-level early warning is triggered; when it is less than the second-level threshold, a second-level early warning is triggered; and when it is less than the third-level threshold, a first-level early warning is triggered. Based on the preset remediation strategy knowledge base, the module matches the corresponding combination of engineering and biological remediation measures to generate structured remediation instructions.

[0043] Preferably, the timestamp alignment operation performed by the data preprocessing module is based on Coordinated Universal Time (UTC). All sensor data are appended with nanosecond-level timestamps at the acquisition end. After being transmitted to the edge computing gateway, they are bucketed and aggregated. The arithmetic mean of the data within the window is taken as the representative value of that time point. Outlier removal adopts an improved 3σ criterion. The mean and standard deviation within a rolling 30-minute window are calculated for each parameter sequence. Data points that exceed the mean plus or minus three times the standard deviation are marked as outliers and removed. Missing value imputation adopts piecewise cubic spline interpolation. The interpolation interval is less than 15 minutes. If it is greater than 15 minutes, it is marked as an invalid period and a backup sensor data source is activated.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] 1. This invention solves the technical deficiency that a single sensor cannot fully reflect the comprehensive ecological status of small wetlands by constructing a multi-source sensor array covering four major media: water, soil, air, and organisms, and realizes full-domain perception of ecological elements.

[0046] 2. A graph attention spatiotemporal fusion network is used to perform cross-media dynamic coupling modeling of multi-source heterogeneous data, which overcomes the limitation of traditional statistical models in being unable to characterize the interaction of complex nonlinear ecological processes and improves the completeness and accuracy of feature representation.

[0047] 3. The prediction model, which introduces a hybrid architecture of long short-term memory network and gated recurrent unit, effectively captures the long-term trends and short-term mutation characteristics of ecological indicators, improves the 24-hour prediction accuracy of key water quality parameters and biodiversity index, and solves the core problem of the lagging early warning of traditional methods. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0049] Figure 2 This is a schematic diagram of the core principle framework of the spatiotemporal feature fusion network based on graph attention mechanism in this invention;

[0050] Figure 3 This is a logical flowchart of the preprocessing and standardization of multi-source heterogeneous ecological sensing data in this invention.

[0051] Figure 4 This is a logical flowchart of the construction of the ecological relationship graph structure and the dynamic edge weight update in this invention;

[0052] Figure 5 This is a schematic diagram of the ecological evolution prediction model based on the hybrid architecture of long short-term memory network and gated recurrent unit in this invention.

[0053] Figure 6This is a schematic diagram of the multi-level interaction relationship and data flow of intelligent early warning and adaptive repair strategy decision-making in this invention. Detailed Implementation

[0054] Please refer to Figures 1 to 6 This invention provides an intelligent monitoring method and system for the ecological restoration of small wetlands, aiming to solve the technical problems in the prior art where a single sensor cannot reflect the comprehensive ecological state, traditional statistical models have low prediction accuracy for complex nonlinear ecological processes, and early warning response is lagging.

[0055] This method deploys a multi-source heterogeneous ecological sensor array to simultaneously collect multi-dimensional dynamic parameters of water bodies, soil, atmosphere, and biological communities, constructing a spatiotemporally aligned high-dimensional ecological state vector. It employs a spatiotemporal feature fusion network based on a graph attention mechanism to perform cross-media coupling modeling of multi-source data, extracting nonlinear interaction characteristics of small wetland ecosystems. A hybrid time-series prediction model, composed of a cascaded long short-term memory network and gated recurrent units, is introduced to perform multi-step advanced predictions of key water quality indicators and biodiversity indices. Combined with a pre-set ecological health threshold system and a dynamic risk assessment rule base, it generates tiered early warning signals and triggers adaptive restoration strategy recommendations, thereby achieving precise perception, advanced early warning, and intelligent intervention of the ecological degradation process of small wetlands.

[0056] The intelligent monitoring method for ecological restoration of small wetlands includes the following steps:

[0057] S1, through a multi-source sensor array deployed in different ecological media of small wetlands, simultaneously acquires water physicochemical parameters, soil environmental parameters, atmospheric meteorological parameters and biological activity parameters, forming a raw ecological perception data stream;

[0058] S2, perform timestamp alignment, outlier removal and missing value imputation on the original ecological perception data stream to generate a standardized multi-dimensional time-series ecological dataset;

[0059] S3. Based on the multidimensional time-series ecological dataset, construct an ecological relationship graph structure with water bodies, soil, atmosphere, and organisms as nodes, where the node features are parameter vectors corresponding to each medium, and the edge weights are determined by the physical and chemical coupling strength between the media.

[0060] S4, input the ecological relationship graph structure into the graph attention spatiotemporal fusion network, calculate the dynamic dependency weights of cross-media features through the multi-head attention mechanism, and output the fused high-order ecological state representation vector;

[0061] S5, the higher-order ecological state characterization vector is input into a hybrid prediction model composed of a long short-term memory network and a gated recurrent unit, to predict dissolved oxygen concentration, ammonia nitrogen content, total phosphorus concentration and benthic biodiversity index in multiple future time steps;

[0062] S6. The prediction results are compared with the preset ecological health classification thresholds. Based on the risk level determination rules, a level 1, level 2 or level 3 ecological early warning signal is generated, and corresponding restoration measures suggestions such as vegetation replanting, aeration and oxygenation or microbial release are output simultaneously.

[0063] In step S1, raw ecological sensing data streams are synchronously acquired through a multi-source sensor array deployed in different ecological media within the micro-wetland. The multi-source sensor array includes a water sensor array, a soil sensor array, an atmospheric sensor array, and a biological activity monitoring unit.

[0064] The water sensor array is installed in water depths of 0.5 to 1.5 meters and is fixed by an underwater support. It includes a dissolved oxygen sensor, an ammonia nitrogen ion selective electrode, a total phosphorus optical detection probe, a turbidity meter, and a multi-parameter water quality analyzer. All sensors are encapsulated in a corrosion-resistant polytetrafluoroethylene shell. The sampling interval is set to 30 seconds to collect seven parameters in real time: water temperature, pH value, dissolved oxygen, conductivity, turbidity, redox potential, and total phosphorus concentration.

[0065] The soil sensor array is buried in the bottom mud layer at depths of 10 cm, 20 cm and 30 cm respectively. It includes a frequency domain reflectance soil moisture sensor, a thermistor soil temperature probe and an electrochemical soil organic matter detection electrode. It is connected to the main control unit through a waterproof junction box and is used to acquire three parameters: soil moisture content, soil temperature and soil organic matter content.

[0066] The atmospheric sensor array is installed at the top of a pole 2 meters above the ground. It includes a digital temperature and humidity sensor, an ultrasonic anemometer and wind vane, and a silicon photodiode illuminance meter, and is used to collect 44 parameters, including air temperature, relative humidity, wind speed and light intensity.

[0067] The biological activity monitoring unit includes a directional microphone array and a passive infrared thermal imaging camera. The microphone sampling rate is 44 kHz, used to calculate the acoustic biological activity index; the thermal imaging resolution is 384×288 pixels, and the frame rate is 9 frames per second, used to extract the infrared thermal imaging biological distribution density.

[0068] All sensors are equipped with nanosecond-level timestamps at the acquisition end, and the data is transmitted to the edge computing gateway via wired or wireless means to form the original ecological perception data stream.

[0069] In step S2, the original ecological sensing data stream undergoes timestamp alignment, outlier removal, and missing value imputation to generate a standardized multidimensional time-series ecological dataset. The timestamp alignment operation is based on Coordinated Universal Time (UTC). All sensor data are bucketed in the edge computing gateway with a 1-minute time window. The arithmetic mean of multiple sampled values ​​of the same parameter within the window is taken as the representative value at that time point, ensuring that all parameters are aligned on the same time base.

[0070] Outlier removal employs a modified 3σ criterion. For each parameter sequence, the mean and standard deviation are calculated within a rolling 30-minute window. If a data point exceeds the range of the mean plus or minus three times the standard deviation within that window, it is marked as an outlier and removed. Missing value imputation uses piecewise cubic spline interpolation, performing interpolation only when the duration of consecutive missing values ​​is less than 15 minutes. If the missing duration is greater than 15 minutes, the data for that period is marked as invalid, and the system automatically switches to a backup sensor data source for imputation.

[0071] After the above processing, a standardized multidimensional time-series ecological dataset with a sampling frequency of once per minute, continuous time span, and no anomalies or missing data is generated. Its data structure is a matrix of time steps × parameter dimensions, with a total of 16 parameter dimensions, corresponding to the aforementioned seven water parameters, three soil parameters, four atmospheric parameters, and two biological parameters.

[0072] In step S3, an ecological relationship graph structure is constructed based on a standardized multidimensional time-series ecological dataset. This graph structure contains four core nodes: water nodes, soil nodes, atmospheric nodes, and biological nodes. The feature vector dimension of the water node is 7, including water temperature, pH value, dissolved oxygen, conductivity, turbidity, redox potential, and total phosphorus concentration; the feature vector dimension of the soil node is 3, including water content, temperature, and organic matter content; the feature vector dimension of the atmospheric node is 4, including air temperature, humidity, wind speed, and light intensity; and the feature vector dimension of the biological node is 2, including acoustic activity index and thermal imaging biological density.

[0073] The connections between nodes are established based on the physicochemical coupling mechanism between ecological media: there is bidirectional material exchange between water and soil, so bidirectional connection edges are set; gas mass transfer between water and atmosphere is mainly through the air-water interface, so unidirectional connection edges from atmosphere to water are set; there is water vapor and heat exchange between soil and atmosphere, so bidirectional connection edges are set; biological activities are mainly affected by the water and soil environment and cause disturbance to water, so unidirectional connection edges from water to organisms, from soil to organisms, and feedback connection edges from organisms to water are set.

[0074] The initial values ​​of the edge weights are derived from long-term observation data. The weight for nitrogen and phosphorus exchange flux between water and soil is set to 0.7, the weight for oxygen mass transfer at the water-air interface is set to 0.5, and the weight for biological disturbance to water is set to 0.3. The remaining edge weights are set according to historical coupling coefficients and are fine-tuned online based on the prediction residual feedback after each prediction cycle. The fine-tuning formula is as follows:

[0075] ;

[0076] Indicates the first Nodes and nodes in each prediction period Edge weights between them Indicates the first Nodes in each prediction period With nodes Edge weights between them The learning rate is 0.01. This is the loss function for the hybrid prediction model.

[0077] In step S4, the ecological relationship graph structure is input into the graph attention spatiotemporal fusion network, which outputs a high-order ecological state representation vector. This network first extracts local neighborhood features through two graph convolutional layers. The graph convolutional layers are implemented using Chebyshev polynomial approximation, with the convolutional kernel order set to 3 and each layer having 64 output channels.

[0078] The first graph convolutional layer receives the original node features and outputs preliminary aggregated neighborhood features. The second graph convolutional layer further expands the receptive field, capturing a wider range of media interaction information. Subsequently, the data enters a multi-head graph attention layer, which contains eight parallel attention heads, each with a query, key, and value projection matrix of dimension 16. For any two nodes... and Their attention score Calculated by scaling the dot product:

[0079] ;

[0080] , For node feature vectors, , For a learnable projection matrix, The dimension is the key vector. Attention weights. Through the Normalization using the softmax function yields:

[0081] ;

[0082] Represents a node The neighborhood group, For two nodes and Attention scores are calculated. The outputs of each attention head are concatenated, linearly transformed, added to the residuals, and finally normalized to generate a high-order ecological state representation vector with a dimension of 128. This vector integrates the dynamic interaction information of four major media: water, soil, air, and organisms, and possesses strong nonlinear expressive power and cross-media coupling characteristics.

[0083] In step S5, the higher-order ecological state representation vector is input into the hybrid time-series prediction model to predict key ecological indicators for the next 24 hours. This model consists of two cascaded long short-term memory network layers and one gated recurrent unit layer. The input is a higher-order ecological state representation vector representing 72 consecutive time steps, corresponding to the historical state over the past 72 hours.

[0084] The first Long Short-Term Memory (LSTM) network layer contains 256 hidden units, which receive the input sequence and output the hidden state sequence; the second LSM network layer also contains 256 hidden units, which receive the hidden states from the first layer and further abstract long-term temporal dependencies.

[0085] The gated recurrent unit layer contains 128 hidden units, receiving the output of the second long short-term memory network layer, integrating long-term trend and short-term fluctuation features. The output layer consists of four fully connected neurons, corresponding to hourly predicted values ​​of dissolved oxygen concentration, ammonia nitrogen content, total phosphorus concentration, and benthic biodiversity index, respectively. A linear activation function is used to ensure that the predicted values ​​are not limited by a range. The model is trained using a weighted loss function. :

[0086] ;

[0087] Mean square error, The mean absolute percentage error is used. An early stopping mechanism is introduced during training: training is terminated when the validation set loss does not decrease for 10 consecutive epochs to prevent overfitting.

[0088] In step S6, the prediction results are compared with a preset ecological health threshold system to generate a graded early warning signal and output restoration suggestions. The ecological health threshold system includes three levels of standards: Level 1 threshold is the critical value for severe damage to ecological function, defined as dissolved oxygen less than 2 mg / L, ammonia nitrogen greater than 2 mg / L, total phosphorus greater than 0.5 mg / L, and benthic biodiversity index less than 1.0; Level 2 threshold is the value for moderate degradation of ecological function, defined as dissolved oxygen less than 4 mg / L, ammonia nitrogen greater than 1.0 mg / L, total phosphorus greater than 0.3 mg / L, and benthic biodiversity index less than 1.5; Level 3 threshold is the value for slight fluctuation in ecological function, defined as dissolved oxygen less than 6 mg / L, ammonia nitrogen greater than 0.5 mg / L, total phosphorus greater than 0.2 mg / L, and benthic biodiversity index less than 2.0.

[0089] The risk level determination rules require that any two of the four indicators simultaneously exceed the same level threshold for more than 3 hours before a corresponding level of warning can be triggered. When a level 3 warning is triggered, the system generates vegetation replanting suggestions, including the types and density of submerged plants such as Vallisneria natans and Myriophyllum spicatum. When a level 2 warning is triggered, the system recommends starting the aeration and oxygenation equipment, setting the start-stop sequence to 10 PM to 6 AM the next day. When a level 1 warning is triggered, the system instructs the release of compound microbial agents at a dosage of 5 grams per cubic meter of water, and simultaneously deploys ecological floating beds at a density of 10 square meters per 100 square meters.

[0090] All remediation measures are derived from a pre-built remediation strategy knowledge base, which contains 12 categories of standard measures. Each category of measures is associated with specific execution parameters, applicable conditions, and expected effects, ensuring the scientific nature and feasibility of the remediation instructions.

[0091] The intelligent monitoring system for small wetland ecological restoration includes a multi-source ecological sensing module, a data preprocessing module, an ecological graph structure construction module, a graph attention spatiotemporal fusion module, a hybrid time-series prediction module, and an intelligent early warning and restoration decision-making module. The multi-source ecological sensing module is responsible for simultaneously collecting 16-dimensional raw parameters of water, soil, atmosphere, and organisms. All sensors are deployed according to media zones and have anti-corrosion, waterproof, and anti-interference capabilities.

[0092] The data preprocessing module performs time alignment, anomaly removal, and missing data imputation, outputting a standardized time-series data matrix updated every minute. The ecosystem graph structure construction module divides the data into four node sets and establishes a dynamic graph structure based on the ecosystem coupling mechanism, with edge weights supporting online updates.

[0093] The graph attention spatiotemporal fusion module generates a 128-dimensional high-order representation vector through graph convolution and multi-head attention mechanisms.

[0094] The hybrid time series forecasting module predicts four key indicators for the next 24 hours based on 72 hours of historical data.

[0095] The intelligent early warning and repair decision-making module generates structured early warning signals and repair instructions based on three-level thresholds and judgment rules, thereby achieving closed-loop intelligent management.

[0096] The system operates within a collaborative architecture between an edge computing gateway and the cloud. The edge computing gateway is responsible for raw data acquisition, preprocessing, and graph structure construction, reducing the load on the cloud. The cloud server deploys a graph attention spatiotemporal fusion network and a hybrid prediction model to complete high-order computation and decision generation. The system performs model fine-tuning at 2:00 AM daily, updating edge weights and network parameters using the residuals between the previous day's actual observations and predictions to ensure the model continuously adapts to the dynamic evolution characteristics of small wetlands.

[0097] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart monitoring method for ecological restoration of small wetlands, characterized in that, include: By simultaneously acquiring water physicochemical parameters, soil environmental parameters, atmospheric meteorological parameters, and biological activity parameters through a multi-source sensor array deployed in different ecological media of small wetlands, a raw ecological perception data stream is formed. The original ecological sensing data stream is processed by timestamp alignment, outlier removal and missing value imputation to generate a standardized multidimensional time-series ecological dataset. Based on the aforementioned multidimensional time-series ecological dataset, an ecological relationship graph structure is constructed with water bodies, soil, atmosphere, and organisms as nodes, including: The standardized multidimensional time-series ecological dataset is divided into four node sets according to media type: water node feature vector dimension is 7, including water temperature, pH value, dissolved oxygen, conductivity, turbidity, redox potential and total phosphorus concentration; soil node feature vector dimension is 3, including water content, temperature and organic matter content; atmosphere node feature vector dimension is 4, including air temperature, humidity, wind speed and light intensity; biological node feature vector dimension is 2, including acoustic activity index and thermal imaging biological density. Based on water-soil exchange flux, air-water interface mass transfer coefficient and biological migration pathway, the connection relationships between nodes are established, including water-soil bidirectional connection, water-atmosphere unidirectional connection, soil-atmosphere bidirectional connection, water-biological unidirectional connection and soil-biological unidirectional connection. The initial values ​​of the edge weights are set based on the historical measured coupling coefficients, and the edge weights are fine-tuned online based on the prediction residual feedback after each prediction cycle. The node features are parameter vectors corresponding to each medium, and the edge weights are determined by the physical and chemical coupling strength between the media. The ecological relationship graph structure is input into a graph attention spatiotemporal fusion network. Dynamic dependency weights of cross-media features are calculated using a multi-head attention mechanism, and the fused high-order ecological state representation vector is output, including: The ecological relationship graph structure is sequentially passed through two graph convolutional layers to extract local neighborhood features. The graph convolutional layers are implemented using Chebyshev polynomial approximation. The output of the second graph convolutional layer is input into a multi-head graph attention layer containing 8 parallel attention heads. The query, key, and value projection matrix of each head has a dimension of 16. Attention scores are calculated by scaling dot products and then normalized by softmax to obtain the attention weights between nodes. The outputs of each attention head are spliced, linearly transformed, residual connected, and layer normalized to generate a high-order ecological state representation vector. The higher-order ecological state representation vector is input into a hybrid prediction model consisting of a long short-term memory network and a gated recurrent unit, to predict dissolved oxygen concentration, ammonia nitrogen content, total phosphorus concentration, and benthic biodiversity index over multiple future time steps.

2. The intelligent monitoring method for ecological restoration of small wetlands according to claim 1, characterized in that, The original ecological sensing data stream is processed by timestamp alignment, outlier removal, and missing value imputation to generate a standardized multidimensional time-series ecological dataset, including: Based on Coordinated Universal Time, all sensor data are bucketed together in the edge computing gateway, and the arithmetic mean of multiple sampled values ​​of the same parameter within the window is taken as the representative value of that time point. For each parameter sequence, the mean and standard deviation are calculated. Data points that exceed the mean plus or minus three times the standard deviation are marked as outliers and removed. For parameter sequences with a continuous missing duration of less than 15 minutes, piecewise cubic spline interpolation is used to fill in the missing values. If the duration is greater than 15 minutes, it is marked as an invalid period and a backup sensor data source is enabled to generate a standardized multidimensional time-series ecological dataset with a sampling frequency of once per minute.

3. The intelligent monitoring method for ecological restoration of small wetlands according to claim 1, characterized in that, The higher-order ecological state representation vector is input into a hybrid prediction model composed of a cascaded long short-term memory network and gated recurrent units to predict dissolved oxygen concentration, ammonia nitrogen content, total phosphorus concentration, and benthic biodiversity index over multiple future time steps, including: The high-order ecological state representation vector of 72 consecutive time steps is used as input, corresponding to the historical state of the past 72 hours. Temporal dependency abstraction is performed sequentially through the first long short-term memory network layer and the second long short-term memory network layer; The output of the second long short-term memory network layer is input into the gated recurrent unit layer to integrate long-term trend and short-term fluctuation characteristics. The output layer, composed of four fully connected neurons, generates hourly predictions of dissolved oxygen, ammonia nitrogen, total phosphorus, and benthic biodiversity index. The activation function of the output layer is a linear function.

4. The intelligent monitoring method for ecological restoration of small wetlands according to claim 3, characterized in that, The training of the hybrid prediction model uses a weighted loss function, which is composed of mean squared error and mean absolute percentage error weighted in a 7:3 ratio, and an early stopping mechanism is introduced to prevent overfitting.

5. The intelligent monitoring method for ecological restoration of small wetlands according to claim 3, characterized in that, The predicted results are compared with preset ecological health grading thresholds, and level 1, 2, or 3 ecological early warning signals are generated according to risk level determination rules. Corresponding remediation measures suggestions are also output simultaneously, including: A Level 3 warning is triggered when any two indicators are predicted to be less than the Level 1 threshold for three consecutive hours; a Level 2 warning is triggered when they are less than the Level 2 threshold; and a Level 1 warning is triggered when they are less than the Level 3 threshold. Based on the warning level, the system matches the corresponding vegetation replanting type, aeration equipment start-up and shutdown sequence, compound microbial agent dosage, and ecological floating bed deployment density from the pre-set restoration strategy knowledge base to generate structured restoration instructions.

6. The intelligent monitoring method for ecological restoration of small wetlands according to claim 5, characterized in that, The repair strategy knowledge base contains 12 The standard measures are categorized, with each category associated with specific implementation parameters, applicable conditions, and expected effects.

7. A smart monitoring system for the ecological restoration of small wetlands, characterized in that, The system for implementing the intelligent monitoring method for ecological restoration of small wetlands as described in any one of claims 1 to 6 comprises: The multi-source ecological sensing module is used to simultaneously collect data on water temperature, pH value, dissolved oxygen, conductivity, turbidity, redox potential, soil moisture content, soil temperature, soil organic matter content, air temperature and humidity, wind speed, light intensity, acoustic biological activity index, and infrared thermal imaging biological distribution density through sensor arrays deployed in small wetland water bodies, bottom sediments, near-ground atmosphere, and vegetation canopy. The data preprocessing module is used to align the raw data output by the multi-source ecological sensing module with a unified time reference, use the sliding window mid-value filtering method to remove abrupt outliers, and use cubic spline interpolation to fill in missing values ​​caused by sampling interruptions, forming a standardized multi-dimensional time series data matrix with a sampling frequency of once per minute. The ecological graph structure construction module is used to divide the standardized multidimensional time series data matrix into four node sets according to the medium type. The feature dimension of each node is the number of parameters contained in the corresponding medium. The connection edges between nodes are established based on the water-soil exchange flux, air-water interface mass transfer coefficient and biological migration path. The initial value of the edge weight is determined by the historical measured coupling coefficient and is dynamically updated during operation. The graph attention spatiotemporal fusion module receives the dynamic graph sequence output by the ecological graph structure construction module, extracts local neighborhood features through two graph convolutional layers, calculates the importance weights between global nodes through a multi-head graph attention layer, and finally aggregates them to generate a high-order ecological state representation vector. The hybrid temporal prediction module inputs the high-order ecological state representation vector into a cascaded structure containing two long short-term memory network layers and one gated recurrent unit layer in time steps. The long short-term memory network layer has 256 hidden units, the gated recurrent unit layer has 128 hidden units, and the output layer has 4 fully connected neurons, corresponding to the hourly prediction values ​​of dissolved oxygen, ammonia nitrogen, total phosphorus, and benthic biodiversity index for the next 24 hours. The intelligent early warning and remediation decision-making module is used to compare the predicted values ​​output by the hybrid time-series prediction module with the three-level ecological health thresholds one by one. When the predicted value of any indicator is less than the first-level threshold for 3 consecutive hours, a third-level early warning is triggered; when it is less than the second-level threshold, a second-level early warning is triggered; and when it is less than the third-level threshold, a first-level early warning is triggered. Based on the preset remediation strategy knowledge base, the module matches the corresponding combination of engineering and biological remediation measures to generate structured remediation instructions.

8. The intelligent monitoring system for ecological restoration of small wetlands according to claim 7, characterized in that, The timestamp alignment operation performed by the data preprocessing module is based on Coordinated Universal Time. All sensor data are appended with nanosecond-level timestamps at the acquisition end and transmitted to the edge computing gateway for bucket aggregation. The arithmetic mean of the data within the window is taken as the representative value of that time point. Outlier removal employs an improved 3σ criterion, calculating the mean and standard deviation within a rolling 30-minute window for each parameter sequence. Data points exceeding the mean plus or minus three times the standard deviation are marked as outliers and removed. Missing value imputation uses piecewise cubic spline interpolation, with the interpolation interval being less than 15 consecutive minutes. If the interval is greater than 15 minutes, it is marked as an invalid period and a backup sensor data source is activated.

Images