Wetland ecosystem degradation identification method and system

Abstract

The invention relates to the technical field of wetland ecosystem degradation identification, in particular to a wetland ecosystem degradation identification method and system. According to the technical scheme, the method comprises the following steps that a comprehensive wetland biodiversity monitoring system is constructed, and multiple types of sensors and high-definition cameras are deployed in a wetland so as to obtain multi-source data related to wetland organisms and environments; preprocessing the obtained original data to enable the original data to be suitable for subsequent artificial intelligence analysis; and biological species identification and biological diversity evaluation are respectively carried out by using an image identification technology based on deep learning and a machine learning algorithm. According to the method, the foundation is laid for analysis by comprehensively and accurately collecting wetland biological and environmental multi-source data and an effective preprocessing method; according to the method, biological species are accurately identified, diversity parameters are evaluated, indexes are reasonably set, trends are reliably analyzed by using a moving average method, data are periodically updated, model parameters and structures are adjusted, wetland changes are self-adapted, and wetland ecology is effectively monitored and evaluated.

Description

Technical Field

[0001] The present invention relates to the technical field of wetland ecosystem degradation identification, and in particular to a wetland ecosystem degradation identification method and system. Background Art

[0002] Wetland ecosystems are one of the most biodiverse and important ecosystems on Earth. However, with the intensification of human activities, wetland ecosystems are facing unprecedented threats, and the degradation of biodiversity is becoming increasingly serious. Accurate and timely identification of the degradation of wetland ecosystems is crucial for wetland protection and management.

[0003] Traditional wetland ecosystem monitoring and assessment methods mainly rely on manual field surveys, such as biodiversity surveys through the same method, mark-recapture method, etc. These methods have many limitations: on the one hand, it is difficult to achieve comprehensive and accurate statistics for some organisms in wetlands that are highly hidden, large in number, and have complex activity patterns (such as small insects, plankton, etc.); on the other hand, comprehensive field surveys of large areas of wetlands require a lot of manpower, material resources and time, and high-frequency monitoring cannot be achieved, making it difficult to capture real-time changes in biodiversity. In addition, biodiversity assessment itself is a complex process that not only considers the number of species, but also the relative abundance, distribution pattern, and interrelationships between species. Existing assessment models often have limitations when processing such complex information and are prone to misjudgment.

[0004] With the development of technology, sensor technology and artificial intelligence technology have provided new ideas for wetland ecosystem monitoring and assessment. Sensors can collect environmental data of wetlands in real time, but the raw data often has problems such as noise and inconsistent data formats, and effective preprocessing is required. At the same time, deep learning and machine learning algorithms have shown strong capabilities in image recognition and data analysis, providing possibilities for biological species identification and biodiversity assessment, but how to effectively integrate these technologies and apply them to the identification of wetland ecosystem degradation still faces challenges.

[0005] In summary, the present application proposes a method and system for identifying wetland ecosystem degradation. Summary of the invention

[0006] The purpose of the present invention is to address the problem of low wetland ecosystem degradation identification accuracy in the background technology and to propose a wetland ecosystem degradation identification method and system.

[0007] In one aspect, the present application provides a method for identifying wetland ecosystem degradation, comprising the following steps:

[0008] Build a comprehensive wetland biodiversity monitoring system, including deploying various types of sensors and high-definition cameras in wetlands to obtain multi-source data related to wetland organisms and the environment;

[0009] Preprocess the acquired raw data to make it suitable for subsequent artificial intelligence analysis;

[0010] Use deep learning-based image recognition technology and machine learning algorithms to identify species and assess biodiversity, respectively, to obtain accurate biodiversity monitoring and assessment results;

[0011] Based on the assessment results, the changing trends of wetland biodiversity are analyzed to promptly detect signs of biodiversity degradation.

[0012] Optionally, sensor types and deployments include:

[0013] Deploy a variety of sensors in different areas of the wetland based on the wetland area, topography, and biological distribution characteristics. Sensor types include but are not limited to water temperature sensors, water quality sensors, soil moisture sensors, and meteorological sensors.

[0014] Camera deployment specifically includes:

[0015] High-definition cameras are installed in different habitats of the wetland, including ground-mounted cameras for monitoring the activities of surface organisms in the wetland, and underwater cameras installed in water bodies.

[0016] Optionally, the preprocessing of the acquired raw data comprises the following steps:

[0017] Data cleaning:

[0018] For the continuous time series data collected by the sensor, the outlier processing method based on the 3σ criterion is adopted. If the data point meets in, is the data mean, σ is the standard deviation, then it is determined to be an outlier and removed;

[0019] For image data, the median filter algorithm is used to remove salt and pepper noise. The formula of the median filter algorithm is:

[0020] y ij =Med{x (i-k),(j-k) ,(k,l∈W)}

[0021] Among them, y ij is the pixel value after filtering, x (i-k),(j-k) is the pixel value within the filter window W;

[0022] Data Standardization:

[0023] All types of data are standardized. For numerical data, the Z-score standardization method is used. The formula is: Among them, x new is the standardized value, x is the original value, μ is the mean, and σ is the standard deviation, so that the data has a uniform mean and standard deviation.

[0024] Optionally, the biological species identification using deep learning-based image recognition technology includes the following steps:

[0025] Image annotation and dataset construction:

[0026] Collect a large amount of wetland biological image data, including images in historical data and newly taken images, covering various types of organisms in wetlands;

[0027] Manually annotate the species and individuals of organisms in the images. The annotation information includes the name, location, and quantity of the organisms. A training dataset of 10,000 images and a test dataset of 2,000 images are constructed.

[0028] Object detection model selection and training:

[0029] The YOLOv5 algorithm is selected as the target detection model. The input of the model is the preprocessed wetland biological image, and the output is the type and number of individuals of the organisms in the image.

[0030] During the training process, the initial learning rate is set to 0.001, the batch size is 16, the number of training rounds is 100, and the cross entropy loss function is used. The formula is:

[0031]

[0032] Where n is the number of categories, y i is the true label, It is the prediction probability. By continuously adjusting the model parameters, the loss function value of the model on the training data set is minimized.

[0033] Optionally, the method of constructing a biodiversity assessment model using a machine learning algorithm comprises the following steps:

[0034] Data collection and compilation:

[0035] Collect data on the types, quantity, and distribution of wetland organisms and related environmental factors, including but not limited to water temperature, water quality parameters, soil moisture, vegetation type, and meteorological data;

[0036] The biological data and environmental factor data were integrated to construct a data set containing 5,000 samples, each of which contained 20 features. The feature weights of the biological data and environmental factor data were preliminarily set according to their potential impact on biodiversity.

[0037] Model selection and training:

[0038] A multi-layer perceptron neural network was selected as the biodiversity assessment model. The number of nodes in the input layer of the model was determined according to the number of input features, the number of nodes in the hidden layer was selected within 50-200, and the number of nodes in the output layer was 1, representing the biodiversity assessment index value.

[0039] The model is trained using the stochastic gradient descent algorithm. The learning rate adopts an adaptive adjustment strategy and the Adagrad algorithm. The update formula of the Adagrad algorithm is:

[0040]

[0041] Among them, w t is the current model parameter, w t+1 is the updated model parameter, η is the initial learning rate, G t is the sum of the squares of the gradients from the initial moment to the moment, ∈ is a small constant, g t is the gradient at the current moment. During the training process, the learning rate and other hyperparameters are adjusted according to the performance of the model on the validation set, so that the model can accurately learn the relationship between biodiversity and environmental factors.

[0042] Optionally, the network structure parameters of the YOLOv5 model include: the convolution kernel size of the convolution layer is 3×3, the step size is 1, and the filling method is same filling; the pooling kernel size of the pooling layer is 2×2, and the step size is 2, so as to improve the model's ability to extract biological image features.

[0043] Optionally, the activation function of the multilayer perceptron adopts a ReLU function, and its formula is:

[0044] f(x)=max(0,x)

[0045] It is used to increase the nonlinear expression ability of the model and improve the model's ability to fit the complex biodiversity-environment relationship.

[0046] Optionally, analyzing the changing trend of wetland biodiversity according to the assessment results includes the following steps:

[0047] Trend analysis indicator settings:

[0048] Set indicators for analyzing biodiversity change trends, including but not limited to the change rate of species richness, the change rate of species evenness, and the change rate of biodiversity index;

[0049] Calculate the value of each indicator at different time points. The time interval is determined according to the monitoring frequency. The indicator value is calculated every 7 days.

[0050] Trend judgment and degradation warning:

[0051] According to the changing trend of the indicator value, the moving average method is used for smoothing, and the moving average window size is 5 to reduce the impact of data fluctuations on trend judgment;

[0052] When the rate of change of the biodiversity index is lower than -0.1 for three consecutive times or the rate of change of species richness and the rate of change of species evenness simultaneously meet specific decline conditions, it is judged that there are signs of biodiversity degradation and an early warning message is issued.

[0053] Optionally, the model update mechanism specifically includes:

[0054] Regularly collect new wetland biodiversity monitoring data and manually verified biodiversity assessment results, with a collection cycle of new data of one month;

[0055] Retrain and update the target detection model and biodiversity assessment model using new data. During the update process, adjust the model's hyperparameters and network structure. Hyperparameters include learning rate and batch size. The network structure includes increasing or decreasing the number of hidden layer nodes.

[0056] System adaptive adjustment:

[0057] Automatically adjust the deployment locations and monitoring parameters of sensors and cameras according to changes in wetland area. For example, when the wetland area expands by more than 10%, increase the number of sensors and cameras in the corresponding area.

[0058] At the same time, data preprocessing methods and model input features are updated according to new biological types and changes in environmental factors.

[0059] On the other hand, the present application provides a wetland ecosystem degradation identification system, comprising:

[0060] Data acquisition module: It is composed of various types of sensors and high-definition cameras. The sensors include water temperature sensors, water quality sensors, soil moisture sensors, and meteorological sensors. They are deployed in different areas of the wetland, such as shallow water areas, deep water areas, swamp areas, and riverbanks, to collect multi-source data related to wetland organisms and the environment.

[0061] Data preprocessing module: including data cleaning unit and data standardization unit. The data cleaning unit uses the outlier processing method based on the 3σ criterion for the continuous time series data collected by the sensor, and uses the median filter algorithm to remove the salt and pepper noise for the image data, and the filter window is set to 3×3; the data standardization unit uses the Z-score standardization method for all numerical data to make the data have a unified mean and standard deviation;

[0062] Biological species identification module: The target detection model is constructed with the YOLOv5 algorithm as the core. The convolution kernel size of the convolution layer is 3×3, the step size is 1, and the padding method is same padding. The pooling kernel size of the pooling layer is 2×2 and the step size is 2. The model input is the preprocessed wetland biological image. The training is performed using a training data set containing 10,000 images and a test data set containing 2,000 images. The initial learning rate is 0.001, the batch size is 16, and the number of training rounds is 100. The cross entropy loss function is used to output the species and number of individuals in the image.

[0063] Biodiversity assessment module: The multi-layer perceptron neural network is used as the model. The number of input layer nodes is determined based on the input containing 5,000 samples and 20 features per sample. The number of hidden layer nodes is selected within 50-200. The activation function uses the ReLU function. The number of output layer nodes is 1, which represents the biodiversity assessment index value. The Adagrad algorithm is used to train the model. The initial learning rate of the Adagrad algorithm is 0.01 and the small constant is 1e-8.

[0064] Biodiversity change trend analysis module: Set the species richness change rate, species evenness change rate, biodiversity index change rate and other analysis indicators, calculate the indicator value every 7 days, use the moving average method to smooth the indicator value change trend, the window size is 5, when the biodiversity index change rate is lower than -0.1 for 3 consecutive times or the species richness change rate and species evenness change rate meet specific decline conditions at the same time, it is determined that biodiversity has signs of degradation and an early warning message is issued;

[0065] System update and adaptation module: Regularly collect new wetland biodiversity monitoring data and manually verified assessment results for retraining and updating target detection models and biodiversity assessment models, adjust sensor and camera deployment locations and monitoring parameters when wetland area changes by more than 10%, and update data preprocessing methods and model input features based on new biological types and changes in environmental factors.

[0066] Compared with the prior art, the present invention has at least one of the following beneficial technical effects:

[0067] 1. The targeted deployment method of the present invention improves the comprehensiveness and accuracy of data collection. Both the physical and chemical environmental data of the wetland and the data of organisms in different habitats can be effectively acquired, providing sufficient and high-quality original data for subsequent analysis. For the continuous time series data collected by the sensor, the outlier processing method based on the 3σ criterion is adopted, which can effectively remove obvious outliers caused by sensor failure or abnormal environmental factors and ensure the reliability of the data. Effectively remove salt and pepper noise and improve image quality. At the same time, it is convenient for the subsequent processing of artificial intelligence algorithms, avoiding the negative impact of data scale differences on model training and analysis.

[0068] 2. Collecting a large amount of image data covering various wetland biological types for processing can enable the model to fully learn the characteristics of wetland biological images during the training process, improve the accuracy of identifying biological species and individual numbers, and realize real-time and continuous monitoring of wetland organisms, especially for birds, large mammals and other organisms that are easily captured by cameras, as well as small organisms.

[0069] 3. Collect data on the types, quantity, and distribution of wetland organisms, as well as data on a variety of related environmental factors. The model can learn the complex relationship between biodiversity and various environmental factors, comprehensively consider factors such as the relative abundance and distribution pattern of species, and improve the accuracy of the assessment. By continuously inputting new data, the model can evaluate changes in biodiversity in real time and promptly discover the degree of biodiversity changes caused by changes in environmental factors.

[0070] 4. Set multi-dimensional analysis indicators such as the change rate of species richness, the change rate of species uniformity, and the change rate of biodiversity index to comprehensively reflect the changes in biodiversity, and be able to promptly and accurately determine signs of biodiversity degradation, providing timely early warning information for wetland protection and management.

[0071] 5. The present invention uses new data to retrain and update the target detection model and biodiversity assessment model, and adjusts the model's hyperparameters and network structure to ensure that the model always adapts to the dynamic changes in wetland biodiversity. The deployment location and monitoring parameters of sensors and cameras are automatically adjusted according to the changes in wetland area, and the data preprocessing method and model input features are updated according to new biological types and environmental factors. This makes the entire monitoring and assessment system highly adaptive and can continue to work accurately in the face of constant changes in wetland ecosystems, ensuring long-term and effective wetland biodiversity monitoring and assessment.

[0072] In summary, the present invention lays the foundation for analysis by comprehensively and accurately collecting multi-source data on wetland organisms and environment and effective preprocessing methods; based on YOLOv5 and multi-layer perceptron combined with appropriate algorithms, parameters accurately identify biological species and evaluate diversity, reasonably set indicators and use moving average method to reliably analyze trends, regularly update data and adjust model parameters and structures, and adapt to wetland changes to effectively monitor and evaluate wetland ecology. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 A flow chart of a wetland ecosystem degradation identification method;

[0074] Figure 2 This is a principle block diagram of a wetland ecosystem degradation identification system. DETAILED DESCRIPTION

[0075] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0076] Example 1

[0077] like Figure 1 As shown, a method for identifying wetland ecosystem degradation is proposed in the present invention, and each step is described in detail below.

[0078] 1. Build a comprehensive wetland biodiversity monitoring system, including deploying various types of sensors and high-definition cameras in wetlands to obtain multi-source data related to wetland organisms and the environment; specifically:

[0079] According to the wetland area, topography and biological distribution characteristics, a variety of sensors are deployed in different areas of the wetland (including shallow water areas, deep water areas, swamp areas, and riverbanks). Sensor types include but are not limited to water temperature sensors, water quality sensors (to detect pH, dissolved oxygen, and nutrient content), soil moisture sensors, and meteorological sensors (to measure temperature, humidity, wind speed, and wind direction);

[0080] Camera deployment specifically includes:

[0081] High-definition cameras are installed in different habitats of wetlands, including ground-mounted cameras for monitoring the activities of surface organisms in wetlands, and underwater cameras installed in water bodies. By deploying various types of sensors (water temperature sensors, water quality sensors, soil moisture sensors, meteorological sensors) and installing high-definition cameras (ground cameras and underwater cameras) in different areas of wetlands (shallow water areas, deep water areas, swamp areas, riverbanks, etc.) according to the area, topography and biological distribution characteristics of wetlands, it is possible to fully cover all key parts of wetland ecosystems and obtain rich multi-source data related to biology and the environment. This targeted deployment method improves the comprehensiveness and accuracy of data collection. Both the physical and chemical environmental data of wetlands and the data of organisms in different habitats can be effectively obtained, providing sufficient and high-quality original data for subsequent analysis.

[0082] 2. Preprocess the acquired raw data to make it suitable for subsequent artificial intelligence analysis; the preprocessing of the acquired raw data includes the following steps:

[0083] Data cleaning:

[0084] For the continuous time series data collected by the sensor, the outlier processing method based on the 3σ criterion is adopted. If the data point meets in, is the data mean, σ is the standard deviation, then it is determined to be an outlier and removed;

[0085] For image data, the median filter algorithm is used to remove salt and pepper noise. The formula of the median filter algorithm is:

[0086] y ij =Med{x (i-k),(j-k) ,(k,l∈W)}

[0087] Among them, y ij is the pixel value after filtering, x (i-k),(j-k) is the pixel value within the filter window W;

[0088] Data Standardization:

[0089] All types of data are standardized. For numerical data, the Z-score standardization method is used. The formula is: Among them, x newis the standardized value, x is the original value, μ is the mean, and σ is the standard deviation, so that the data has a unified mean and standard deviation, which is convenient for subsequent analysis. For the continuous time series data collected by the sensor, the outlier processing method based on the 3σ criterion is used, which can effectively remove the obvious outliers caused by sensor failure or abnormal environmental factors, and ensure the reliability of the data. For image data, the median filter algorithm (filter window is set to 3×3) can be used to effectively remove salt and pepper noise and improve image quality. At the same time, the Z-score standardization method is used for all numerical data, so that data of different types and magnitudes have a unified mean and standard deviation, which is convenient for the subsequent processing of artificial intelligence algorithms and avoids the negative impact of data scale differences on model training and analysis.

[0090] 3. Use deep learning-based image recognition technology and machine learning algorithms to identify species and assess biodiversity, respectively, to obtain accurate biodiversity monitoring and assessment results; using deep learning-based image recognition technology to identify species includes the following steps:

[0091] Image annotation and dataset construction:

[0092] Collect a large amount of wetland biological image data, including images in historical data and newly taken images, covering various types of organisms in wetlands (such as birds, large mammals, small insects, and plankton);

[0093] Manually annotate the species and individuals of organisms in the images. The annotation information includes the name, location, and quantity of the organisms. A training dataset of 10,000 images and a test dataset of 2,000 images are constructed.

[0094] Object detection model selection and training:

[0095] The YOLOv5 algorithm is selected as the target detection model. The input of the model is the preprocessed wetland biological image, and the output is the type and number of individuals of the organisms in the image.

[0096] During the training process, the initial learning rate is set to 0.001, the batch size is 16, the number of training rounds is 100, and the cross entropy loss function is used. The formula is:

[0097]

[0098] Where n is the number of categories, t i is the true label, It is the prediction probability. By continuously adjusting the model parameters, the loss function value of the model on the training data set is minimized.

[0099] Building a biodiversity assessment model using machine learning algorithms involves the following steps:

[0100] Data collection and compilation:

[0101] Collect data on the types, quantity, and distribution of wetland organisms and related environmental factor data, including but not limited to water temperature, water quality parameters, soil moisture, vegetation type, meteorological data, etc.;

[0102] The biological data and environmental factor data were integrated to construct a data set containing 5,000 samples, each of which contained 20 features. The feature weights of the biological data and environmental factor data were preliminarily set according to their potential impact on biodiversity.

[0103] Model selection and training:

[0104] A multi-layer perceptron neural network was selected as the biodiversity assessment model. The number of nodes in the input layer of the model was determined according to the number of input features, the number of nodes in the hidden layer was selected within 50-200, and the number of nodes in the output layer was 1, representing the biodiversity assessment index value.

[0105] The model is trained using the stochastic gradient descent algorithm. The learning rate adopts an adaptive adjustment strategy and the Adagrad algorithm. The update formula of the Adagrad algorithm is:

[0106]

[0107] Among them, w t is the current model parameter, w t+1 is the updated model parameter, η is the initial learning rate, G t is the sum of the squares of the gradients from the initial moment to the moment, ∈ is a small constant, g t is the gradient at the current moment. During the training process, the learning rate and other hyperparameters are adjusted according to the performance of the model on the validation set, so that the model can accurately learn the relationship between biodiversity and environmental factors.

[0108] The network structure parameters of the YOLOv5 model include: the convolution kernel size of the convolution layer is 3×3, the step size is 1, and the padding method is same padding; the pooling kernel size of the pooling layer is 2×2, and the step size is 2, in order to improve the model's ability to extract biological image features.

[0109] The activation function of the multi-layer perceptron uses the ReLU function, and its formula is:

[0110] f(x)=max(0,x)

[0111] It is used to increase the nonlinear expression ability of the model and improve the model's ability to fit the complex biodiversity-environment relationship.

[0112] A large amount of image data covering various wetland biological types (including historical data and newly taken images) was collected and annotated in detail, and a training data set of 10,000 images and a test data set of 2,000 images were constructed, providing rich and representative samples for model training. The YOLOv5 algorithm was selected as the target detection model, and its network structure parameters (convolutional layer convolution kernel size of 3×3, step size of 1, padding method of same padding, pooling layer pooling kernel size of 2×2, step size of 2) were reasonably set. Combined with appropriate training parameters (initial learning rate of 0.001, batch size of 16, number of training rounds of 100, cross entropy loss function), the model can fully learn the characteristics of wetland biological images during training, improve the accuracy of biological species and individual number recognition, and realize real-time and continuous monitoring of wetland biological species, especially for birds, large mammals and other creatures that are easily captured by cameras, as well as small creatures (using underwater cameras combined with image recognition technology) can be effectively monitored.

[0113] The types, quantity, and distribution data of wetland organisms and various related environmental factor data (water temperature, water quality parameters, soil moisture, vegetation type, meteorological data, etc.) were collected to construct a data set containing 5,000 samples and 20 features for each sample. The feature weights were preliminarily set according to the potential impact of biological data and environmental factor data on biodiversity, providing comprehensive information for the biodiversity assessment model. The multi-layer perceptron neural network was selected as the model, and the number of input layer nodes and hidden layer nodes (50-200) were reasonably set. The ReLU activation function was used to increase the nonlinear expression ability of the model, and the Adagrad algorithm (initial learning rate was 0.01, and the small constant was 1e-8) was used to train the model, so that the model can learn the complex relationship between biodiversity and various environmental factors, comprehensively consider the relative abundance and distribution pattern of species, and improve the accuracy of the assessment. By continuously inputting new data, the model can evaluate the changes in biodiversity in real time and promptly discover the degree of biodiversity changes caused by changes in environmental factors (such as water level changes and water temperature changes affecting biological habitats).

[0114] 4. Based on the assessment results, analyze the changing trends of wetland biodiversity and promptly identify signs of biodiversity degradation.

[0115] Analyzing the changing trend of wetland biodiversity based on the assessment results includes the following steps:

[0116] Trend analysis indicator settings:

[0117] Set indicators for analyzing biodiversity change trends, including but not limited to the change rate of species richness, the change rate of species evenness, and the change rate of biodiversity index;

[0118] Calculate the value of each indicator at different time points. The time interval is determined according to the monitoring frequency. The indicator value is calculated every 7 days.

[0119] Trend judgment and degradation warning:

[0120] According to the changing trend of the indicator value, the moving average method is used for smoothing, and the moving average window size is 5 to reduce the impact of data fluctuations on trend judgment;

[0121] When the biodiversity index change rate is lower than -0.1 for three consecutive times or the species richness change rate and species evenness change rate meet specific decline conditions at the same time, it is judged that biodiversity shows signs of degradation and an early warning message is issued. Multi-dimensional analysis indicators such as species richness change rate, species evenness change rate, and biodiversity index change rate are set, and the indicator values ​​are calculated every 7 days to fully reflect the changes in biodiversity. The moving average method (window size is 5) is used to smooth the trend of indicator value changes, which effectively reduces the impact of data fluctuations on trend judgment and improves the reliability of trend analysis. By setting reasonable degradation judgment conditions (such as the biodiversity index change rate is lower than -0.1 for three consecutive times or the species richness change rate and species evenness change rate meet specific decline conditions at the same time), it is possible to timely and accurately judge the signs of biodiversity degradation and provide timely early warning information for wetland protection and management.

[0122] The model update mechanism specifically includes:

[0123] Regularly collect new wetland biodiversity monitoring data and manually verified biodiversity assessment results, with a collection cycle of new data of one month;

[0124] Retrain and update the target detection model and biodiversity assessment model using new data. During the update process, adjust the model's hyperparameters and network structure. Hyperparameters include learning rate and batch size. The network structure includes increasing or decreasing the number of hidden layer nodes.

[0125] System adaptive adjustment:

[0126] Automatically adjust the deployment locations and monitoring parameters of sensors and cameras according to changes in wetland area. For example, when the wetland area expands by more than 10%, increase the number of sensors and cameras in the corresponding area.

[0127] At the same time, according to the new biological types and changes in environmental factors, update the data preprocessing methods and model input features to ensure that the entire monitoring and assessment system can always accurately reflect the actual situation of wetland biodiversity. Establish a mechanism to regularly (1 month) collect new wetland biodiversity monitoring data and manually verified assessment results, use the new data to retrain and update the target detection model and biodiversity assessment model, and adjust the model's hyperparameters (learning rate, batch size) and network structure (increase or decrease the number of hidden layer nodes) to ensure that the model always adapts to the dynamic changes of wetland biodiversity. Automatically adjust the deployment location and monitoring parameters of sensors and cameras according to changes in wetland area (more than 10%), and update data preprocessing methods and model input features according to new biological types and changes in environmental factors, so that the entire monitoring and assessment system has a strong adaptive ability, can continue to work accurately in the face of constant changes in wetland ecosystems, and ensure long-term and effective wetland biodiversity monitoring and assessment.

[0128] Example 2

[0129] like Figure 2 As shown, this embodiment provides a wetland ecosystem degradation identification system, including:

[0130] Data acquisition module: It is composed of various types of sensors and high-definition cameras. The sensors include water temperature sensors, water quality sensors, soil moisture sensors, and meteorological sensors, which are deployed in different areas of the wetland, such as shallow water areas, deep water areas, swamp areas, and riverbanks. The accuracy of the water temperature sensor reaches ±0.2℃, the accuracy of the water quality sensor for pH reaches ±0.02pH units, and the accuracy of dissolved oxygen reaches ±0.2mg / L. The measurement error of the soil moisture sensor is less than 5%, and the meteorological sensor meets the relevant standards of the World Meteorological Organization; the cameras include ground cameras and underwater cameras. The resolution of the ground camera is not less than 1920×1080 pixels, the frame rate is not less than 25 frames / second, and the field of view is 30°-90°. The waterproof depth of the underwater camera is not less than 5 meters and the resolution is not less than 1280×720 pixels. It is used to collect multi-source data related to wetland organisms and the environment;

[0131] Data preprocessing module: including data cleaning unit and data standardization unit. The data cleaning unit uses the outlier processing method based on the 3σ criterion for the continuous time series data collected by the sensor, and uses the median filter algorithm (the filter window can be set to 3×3) to remove the salt and pepper noise for the image data; the data standardization unit uses the Z-score standardization method for all numerical data to make the data have a unified mean and standard deviation;

[0132] Biological species identification module: The target detection model is constructed with the YOLOv5 algorithm as the core. The convolution kernel size of the convolution layer is 3×3, the step size is 1, and the padding method is same padding. The pooling kernel size of the pooling layer is 2×2 and the step size is 2. The model input is the preprocessed wetland biological image. The training is performed using a training data set containing 10,000 images and a test data set containing 2,000 images. The initial learning rate is 0.001, the batch size is 16, and the number of training rounds is 100. The cross entropy loss function is used to output the species and number of individuals in the image.

[0133] Biodiversity assessment module: The multi-layer perceptron neural network is used as the model. The number of input layer nodes is determined based on the input of 5,000 samples with 20 features per sample (integrating biological data and environmental factor data). The number of hidden layer nodes is selected within 50-200. The activation function uses the ReLU function. The number of output layer nodes is 1, which represents the biodiversity assessment index value. The Adagrad algorithm (initial learning rate is 0.01, and the small constant is 1e-8) is used to train the model.

[0134] Biodiversity change trend analysis module: Set the species richness change rate, species evenness change rate, biodiversity index change rate and other analysis indicators, calculate the indicator value every 7 days, and use the moving average method (window size is 5) to smooth the indicator value change trend. When the biodiversity index change rate is lower than -0.1 for three consecutive times or the species richness change rate and species evenness change rate meet specific decline conditions at the same time, it is determined that biodiversity has signs of degradation and an early warning message is issued;

[0135] System update and adaptation module: Regularly (one-month cycle) collect new wetland biodiversity monitoring data and manually verified assessment results for retraining and updating the target detection model and biodiversity assessment model (including adjusting hyperparameters and network structure such as learning rate, batch size, number of hidden layer nodes, etc.), adjust the sensor and camera deployment locations and monitoring parameters according to changes in wetland area (more than 10%), and update data preprocessing methods and model input features based on new biological types and changes in environmental factors.

[0136] The above specific embodiments are only several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A method for identifying wetland ecosystem degradation, characterized in that: The following steps are involved: Build a comprehensive wetland biodiversity monitoring system, including deploying various types of sensors and high-definition cameras in wetlands to obtain multi-source data related to wetland organisms and the environment; Preprocess the acquired raw data to make it suitable for subsequent artificial intelligence analysis; Use deep learning-based image recognition technology and machine learning algorithms to identify species and assess biodiversity, respectively, to obtain accurate biodiversity monitoring and assessment results; Based on the assessment results, the changing trends of wetland biodiversity are analyzed to promptly detect signs of biodiversity degradation.

2. A wetland ecosystem degradation identification method according to claim 1, characterized in that: Sensor types and deployment include: Deploy a variety of sensors in different areas of the wetland based on the wetland area, topography, and biological distribution characteristics. Sensor types include but are not limited to water temperature sensors, water quality sensors, soil moisture sensors, and meteorological sensors. Camera deployment specifically includes: High-definition cameras are installed in different habitats of the wetland, including ground-mounted cameras for monitoring the activities of surface organisms in the wetland, and underwater cameras installed in water bodies.

3. A wetland ecosystem degradation identification method according to claim 1, characterized in that: The preprocessing of the acquired raw data comprises the following steps: Data cleaning: For the continuous time series data collected by the sensor, the outlier processing method based on the 3σ criterion is adopted. If the data point meets in, is the data mean, σ is the standard deviation, then it is determined to be an outlier and removed; For image data, the median filter algorithm is used to remove salt and pepper noise. The formula of the median filter algorithm is: y ij =With{x (i-k),(j-k) ,(k,l∈W)} Among them, y ij is the pixel value after filtering, x (i-k),(j-k) is the pixel value within the filter window W; Data Standardization: All types of data are standardized. For numerical data, the Z-score standardization method is used. The formula is: Among them, x new is the standardized value, x is the original value, μ is the mean, and σ is the standard deviation, so that the data has a uniform mean and standard deviation.

4. A wetland ecosystem degradation identification method according to claim 1, characterized in that: The method of using deep learning-based image recognition technology to identify biological species includes the following steps: Image annotation and dataset construction: Collect a large amount of wetland biological image data, including images in historical data and newly taken images, covering various types of organisms in wetlands; Manually annotate the species and individuals of organisms in the images. The annotation information includes the name, location, and quantity of the organisms. A training dataset of 10,000 images and a test dataset of 2,000 images are constructed. Object detection model selection and training: The YOLOv5 algorithm is selected as the target detection model. The input of the model is the preprocessed wetland biological image, and the output is the type and number of individuals of the organisms in the image. During the training process, the initial learning rate is set to 0.001, the batch size is 16, the number of training rounds is 100, and the cross entropy loss function is used. The formula is: Where n is the number of categories, y i is the true label, It is the prediction probability. By continuously adjusting the model parameters, the loss function value of the model on the training data set is minimized.

5. A wetland ecosystem degradation identification method according to claim 1, characterized in that: The method of constructing a biodiversity assessment model using a machine learning algorithm comprises the following steps: Data collection and compilation: Collect data on the types, quantity, and distribution of wetland organisms and related environmental factors, including but not limited to water temperature, water quality parameters, soil moisture, vegetation type, and meteorological data; The biological data and environmental factor data were integrated to construct a data set containing 5,000 samples, each of which contained 20 features. The feature weights of the biological data and environmental factor data were preliminarily set according to their potential impact on biodiversity. Model selection and training: A multi-layer perceptron neural network was selected as the biodiversity assessment model. The number of nodes in the input layer of the model was determined according to the number of input features, the number of nodes in the hidden layer was selected within 50-200, and the number of nodes in the output layer was 1, representing the biodiversity assessment index value. The model is trained using the stochastic gradient descent algorithm. The learning rate adopts an adaptive adjustment strategy and the Adagrad algorithm. The update formula of the Adagrad algorithm is: Among them, w t is the current model parameter, w t+1 is the updated model parameter, η is the initial learning rate, G t is the sum of squared gradients from the initial moment to the moment, ∈ is a small constant, g t is the gradient at the current moment. During the training process, the learning rate and other hyperparameters are adjusted according to the performance of the model on the validation set, so that the model can accurately learn the relationship between biodiversity and environmental factors.

6. A wetland ecosystem degradation identification method according to claim 4, characterized in that: The network structure parameters of the YOLOv5 model include: the convolution kernel size of the convolution layer is 3×3, the step size is 1, and the padding method is same padding; the pooling kernel size of the pooling layer is 2×2, and the step size is 2, so as to improve the model's ability to extract biological image features.

7. A wetland ecosystem degradation identification method according to claim 5, characterized in that: The activation function of the multilayer perceptron adopts the ReLU function, and its formula is: f(x)=max(0,x) It is used to increase the nonlinear expression ability of the model and improve the model's ability to fit the complex biodiversity-environment relationship.

8. A wetland ecosystem degradation identification method according to claim 1, characterized in that: The analysis of the changing trend of wetland biodiversity based on the assessment results includes the following steps: Trend analysis indicator settings: Set indicators for analyzing biodiversity change trends, including but not limited to the change rate of species richness, the change rate of species evenness, and the change rate of biodiversity index; Calculate the value of each indicator at different time points. The time interval is determined according to the monitoring frequency. The indicator value is calculated every 7 days. Trend judgment and degradation warning: According to the changing trend of the indicator value, the moving average method is used for smoothing, and the moving average window size is 5 to reduce the impact of data fluctuations on trend judgment; When the rate of change of the biodiversity index is lower than -0.1 for three consecutive times or the rate of change of species richness and the rate of change of species evenness simultaneously meet specific decline conditions, it is judged that there are signs of biodiversity degradation and an early warning message is issued.

9. A wetland ecosystem degradation identification method according to claim 1, characterized in that: The model update mechanism specifically includes: Regularly collect new wetland biodiversity monitoring data and manually verified biodiversity assessment results, with a collection cycle of new data of one month; Retrain and update the target detection model and biodiversity assessment model using new data. During the update process, adjust the model's hyperparameters and network structure. Hyperparameters include learning rate and batch size. The network structure includes increasing or decreasing the number of hidden layer nodes. System adaptive adjustment: Automatically adjust the deployment locations and monitoring parameters of sensors and cameras according to changes in wetland area. For example, when the wetland area expands by more than 10%, increase the number of sensors and cameras in the corresponding area. At the same time, data preprocessing methods and model input features are updated according to new biological types and changes in environmental factors.

10. A wetland ecosystem degradation identification system, comprising: Data acquisition module: It is composed of various types of sensors and high-definition cameras. The sensors include water temperature sensors, water quality sensors, soil moisture sensors, and meteorological sensors. They are deployed in different areas of the wetland, such as shallow water areas, deep water areas, swamp areas, and riverbanks, to collect multi-source data related to wetland organisms and the environment. Data preprocessing module: including data cleaning unit and data standardization unit. The data cleaning unit uses the outlier processing method based on the 3σ criterion for the continuous time series data collected by the sensor, and uses the median filter algorithm to remove the salt and pepper noise for the image data, and the filter window is set to 3×3; the data standardization unit uses the Z-score standardization method for all numerical data to make the data have a unified mean and standard deviation; Biological species identification module: The target detection model is constructed with the YOLOv5 algorithm as the core. The convolution kernel size of the convolution layer is 3×3, the step size is 1, and the padding method is same padding. The pooling kernel size of the pooling layer is 2×2 and the step size is 2. The model input is the preprocessed wetland biological image. The training is performed using a training data set containing 10,000 images and a test data set containing 2,000 images. The initial learning rate is 0.001, the batch size is 16, and the number of training rounds is 100. The cross entropy loss function is used to output the species and number of individuals in the image. Biodiversity assessment module: The multi-layer perceptron neural network is used as the model. The number of input layer nodes is determined based on the input containing 5,000 samples and 20 features per sample. The number of hidden layer nodes is selected within 50-200. The activation function uses the ReLU function. The number of output layer nodes is 1, which represents the biodiversity assessment index value. The Adagrad algorithm is used to train the model. The initial learning rate of the Adagrad algorithm is 0.01 and the small constant is 1e-8. Biodiversity change trend analysis module: Set the species richness change rate, species evenness change rate, biodiversity index change rate and other analysis indicators, calculate the indicator value every 7 days, use the moving average method to smooth the indicator value change trend, the window size is 5, when the biodiversity index change rate is lower than -0.1 for 3 consecutive times or the species richness change rate and species evenness change rate meet specific decline conditions at the same time, it is determined that biodiversity has signs of degradation and an early warning message is issued; System update and adaptation module: Regularly collect new wetland biodiversity monitoring data and manually verified assessment results for retraining and updating target detection models and biodiversity assessment models, adjust sensor and camera deployment locations and monitoring parameters when wetland area changes by more than 10%, and update data preprocessing methods and model input features based on new biological types and changes in environmental factors.

Images