Coastal wetland intelligent monitoring method and system based on artificial intelligence

Abstract

The invention relates to the technical field of ecological environment monitoring, and discloses a coastal wetland intelligent monitoring method and system based on artificial intelligence, and the method comprises the steps: collecting unmanned plane data, satellite remote sensing data, Internet of Things sensor data and water quality monitoring buoy data of a coastal wetland; the method comprises the following steps: processing satellite remote sensing data by adopting a wavelet threshold denoising algorithm based on an attention mechanism, calibrating Internet of Things sensor data by adopting an LSTM network, and carrying out data space-time alignment based on a space-time attention fusion model to obtain preprocessed data; inputting the preprocessed data into a Transform-ResNet hybrid model to carry out environmental change evaluation, and outputting an ecological health index; when the predicted ecological health index is lower than a threshold value, a PPO algorithm is adopted to dynamically adjust a monitoring strategy according to the early warning level, and an early warning report is pushed; the whole process is intelligent, manual intervention is greatly reduced, and support is provided for coastal wetland ecological protection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ecological environment monitoring, in particular to a coastal wetland intelligent monitoring method and system based on artificial intelligence. BACKGROUND

[0002] As an important ecological system at the land-sea interface, coastal wetlands are characterized by rich biodiversity and outstanding ecological functions, making them a key area for global climate change research and ecological protection. However, traditional coastal wetland monitoring methods have the following problems: (1) relying on manual field surveys, which is inefficient and costly; (2) single remote sensing data sources cannot fully reflect the ecological characteristics of wetlands; (3) the lack of intelligence in monitoring data processing makes it difficult to achieve dynamic and accurate ecological assessment; (4) the ability to predict wetland changes is insufficient, and it is difficult to timely warn ecological risks. SUMMARY

[0003] The present application aims to solve the above problems and provides a coastal wetland intelligent monitoring method and system based on artificial intelligence.

[0004] The first aspect of the present application provides a coastal wetland intelligent monitoring method based on artificial intelligence, which comprises the following steps: Collecting unmanned aerial vehicle data, satellite remote sensing data, Internet of Things sensor data, and water quality monitoring buoy data of the coastal wetland; Processing the satellite remote sensing data using a wavelet threshold denoising algorithm based on attention mechanism, and performing data spatio-temporal alignment based on a spatio-temporal attention fusion model to obtain preprocessed data; Inputting the preprocessed data into a Transformer-ResNet hybrid model for environmental change assessment, and outputting an ecological health index, wherein the Transformer-ResNet hybrid model includes an adaptive multi-scale spatio-temporal attention module, a multi-task learning branch, and a dynamic weight adjustment mechanism; When the predicted ecological health index is lower than a threshold value, using a PPO algorithm to dynamically adjust the monitoring strategy according to the warning level and push a warning report.

[0005] Optionally, in the first implementation manner of the first aspect of the present application, the satellite remote sensing data is processed using a wavelet threshold denoising algorithm based on attention mechanism, which comprises: Selecting a db4 wavelet basis as a decomposition basis function to perform multi-scale wavelet decomposition on the satellite remote sensing data to obtain low-frequency approximation coefficients and high-frequency detail coefficients at different scales; Inputting the high-frequency detail coefficients into an attention calculation network, mapping the features through a fully connected layer, and then normalizing the features through a Softmax function to obtain the attention weights corresponding to each high-frequency detail coefficient; The high-frequency detail coefficient is thresholded based on the attention weight, and the processed high-frequency detail coefficient and the original low-frequency approximation coefficient are reconstructed by wavelet to obtain the denoised satellite remote sensing data.

[0006] Optionally, in the second implementation manner of the first aspect of the present application, the data spatio-temporal alignment based on the spatio-temporal attention fusion model to obtain the preprocessed data comprises: The monitoring area of the coastal wetland is divided into a plurality of spatial grids in the spatial dimension and a plurality of 1-hour time segments in the time dimension to form a grid system covering the whole space-time of the monitoring area; The processed unmanned aerial vehicle data, satellite remote sensing data, Internet of Things sensor data and water quality monitoring buoy data are subjected to spatio-temporal feature extraction; The spatio-temporal features are input into the spatio-temporal attention calculation module, classified according to the time segments, and in each time segment, the correlation of the spatio-temporal features with the grid system is calculated, and the attention weight of the spatio-temporal features in the corresponding grid is calculated through the full connection layer and the Softmax function according to the correlation size; The attention weights are weighted and summed to obtain the comprehensive data of the space-time grid in the corresponding time segment, and all space-time grids and time segments are traversed to form the preprocessed data.

[0007] Optionally, in the third implementation manner of the first aspect of the present application, the preprocessed data is input into the Transformer-ResNet hybrid model for environment change evaluation, and the ecological health index is output, comprising: The preprocessed data after spatio-temporal alignment is input into the Transformer-ResNet hybrid model, the data is subjected to spatial feature extraction through the convolution layer of the ResNet network, and the data is converted into a feature vector conforming to the processing format of the Transformer through the embedding layer of the Transformer network; The spatial features extracted by the ResNet and the feature vectors embedded by the Transformer are input into the adaptive multi-scale spatio-temporal attention module, and the spatio-temporal features of different scales are automatically learned through the multi-scale feature extraction and adaptive weight distribution mechanism to obtain multi-scale spatial attention weights; The features after attention adjustment are input into the multi-task learning branch, and each branch adopts a full connection layer to map the features to obtain the numerical values of the four environment evaluation indexes; Based on the wetland type, seasonal change and historical ecological data, the weights of each environment evaluation are automatically adjusted, the prediction results of the multi-task learning branches are fused by weighting to generate the final ecological health index.

[0008] Optionally, in a fourth implementation form of the first aspect of the present application, the multi-task learning branch comprises a vegetation coverage evaluation branch, a water quality compliance rate evaluation branch, a soil salinization degree evaluation branch, and a hydrological connectivity evaluation branch.

[0009] Optionally, in a fifth implementation form of the first aspect of the present application, when the predicted ecological health index is lower than the threshold value, the PPO algorithm is used to dynamically adjust the monitoring strategy according to the warning level and push the warning report, which comprises: When the predicted ecological health index is lower than the threshold value, the current warning level and the corresponding ecological health index value are recorded as the initial input information of the PPO algorithm; The policy network and the value network of the PPO algorithm are initialized, and a plurality of groups of candidate monitoring strategies are generated according to the current warning level and the initialized policy network; Each group of candidate strategies is applied to the simulated monitoring scene, the environmental state changes during the execution of the strategy are recorded, and the performance score of each group of strategies is evaluated by the value network; According to the performance score of the candidate strategy, the advantage function of the strategy is calculated, and the parameter of the strategy network is updated by the clipping objective function of the PPO algorithm; After a plurality of iterations, the optimal monitoring strategy under the current warning level is determined, and the warning report is generated and pushed according to the finally determined warning level and the optimal monitoring strategy.

[0010] Optionally, in a sixth implementation form of the first aspect of the present application, if the ecological health index is between 0.5 and 0.6, it is determined as yellow warning; if the ecological health index is between 0.4 and 0.5, it is determined as orange warning; and if the ecological health index is lower than 0.4, it is determined as red warning.

[0011] The second aspect of the present application provides an intelligent monitoring system for coastal wetlands based on artificial intelligence, which comprises: A collection module for collecting unmanned aerial vehicle data, satellite remote sensing data, Internet of Things sensor data and water quality monitoring buoy data of the coastal wetland; A preprocessing module for processing satellite remote sensing data by using a wavelet threshold denoising algorithm based on an attention mechanism, and performing data space-time alignment based on a space-time attention fusion model to obtain preprocessed data; An evaluation module for inputting the preprocessed data into a Transformer-ResNet hybrid model to evaluate environmental changes and output an ecological health index, wherein the Transformer-ResNet hybrid model comprises an adaptive multi-scale space-time attention module, a multi-task learning branch and a dynamic weight adjustment mechanism; The monitoring module is configured to dynamically adjust the monitoring strategy according to the warning level and push a warning report when the predicted ecological health index is lower than a threshold value.

[0012] The third aspect of the present application provides an artificial intelligence-based coastal wetland intelligent monitoring device, which comprises a memory and at least one processor, and the memory stores instructions; the at least one processor invokes the instructions in the memory to enable the artificial intelligence-based coastal wetland intelligent monitoring device to perform the steps of the artificial intelligence-based coastal wetland intelligent monitoring method according to any one of the above.

[0013] The fourth aspect of the present application provides a computer-readable storage medium, which stores instructions, and the instructions are executed by a processor to implement the steps of the artificial intelligence-based coastal wetland intelligent monitoring method according to any one of the above.

[0014] In the technical solution provided by the present application, the unmanned aerial vehicle data, satellite remote sensing data, Internet of Things sensor data and water quality monitoring buoy data of the coastal wetland are collected; the satellite remote sensing data is processed by using a wavelet threshold denoising algorithm based on an attention mechanism, the data is spatio-temporally aligned based on a spatio-temporal attention fusion model, and preprocessed data is obtained; the preprocessed data is input into a Transformer-ResNet hybrid model for environmental change evaluation, and an ecological health index is output; when the predicted ecological health index is lower than a threshold value, a PPO algorithm is used to dynamically adjust the monitoring strategy according to the warning level and push a warning report; the present application can comprehensively cover multi-dimensional ecological factors such as wetland vegetation, topography, soil and water body, avoiding the limitation that traditional single data source can only reflect local information; by using the Transformer-ResNet hybrid model comprising an adaptive multi-scale spatio-temporal attention module, a multi-task learning branch and a dynamic weight adjustment mechanism, the defects of traditional evaluation dimension being single and weight being fixed can be overcome, the precise and dynamic determination of the wetland ecological health state can be realized, and an ecological health index that conforms to the actual situation can be output; the PPO algorithm can automatically determine the warning level and dynamically select the adaptive monitoring strategy, avoiding the problems of traditional artificial response lag and poor strategy adaptability, and gaining time for risk prevention and control; the full-process intelligentization greatly reduces artificial intervention, reduces labor cost and operation complexity, has strong practicability and generalizability, and provides efficient and universal technical support for coastal wetland ecological protection. BRIEF DESCRIPTION OF DRAWINGS

[0015] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments with reference made to the accompanying drawings. The drawings are for purposes of illustration only and are not considered as limiting the application.

[0016] Figure 1 A flowchart of the intelligent monitoring method of the coastal wetland based on artificial intelligence provided by the embodiment of the present application is shown in FIG. 1. Figure 2 A structural schematic diagram of the intelligent monitoring system of the coastal wetland based on artificial intelligence provided by the embodiment of the present application is shown in FIG. 2. Figure 3 A structural schematic diagram of the intelligent monitoring device of the coastal wetland based on artificial intelligence provided by the embodiment of the present application is shown in FIG. 3. DETAILED DESCRIPTION

[0017] The terms "first", "second", "third", "fourth" and the like in the description, claims, and drawings of the present application, and those above (if any) are used for distinguishing between similar objects, and do not necessarily have to follow a specific order or sequence. It is to be understood that the use of the terms herein is merely for distinguishing between the objects described, and does not necessarily indicate a required or particular chronology or sequence of events, unless specified otherwise. It is to be understood that the data used herein can be interchanged, under appropriate circumstances, unless otherwise expressly specified and / or except where inapplicable. Furthermore, the term "comprising" or "containing" and variations thereof as used herein is used generically and inclusively, in that it specifies the presence of stated features, integers, steps or components, but does not preclude the presence or addition of one or more other features, integers, steps, components or groups thereof that are not expressly stated, but which are present or can be added.

[0018] For the convenience of understanding, the specific flow of the embodiment of the present application is described below, please refer to Figure 1 The flowchart of the intelligent monitoring method of the coastal wetland based on artificial intelligence provided by the embodiment of the present application, which specifically includes the following steps: Step 101, collecting unmanned aerial vehicle data, satellite remote sensing data, Internet of Things sensor data and water quality monitoring buoy data of the coastal wetland; In this embodiment, in the data collection link of the coastal wetland, a space-air-ground-water four-dimensional cooperative collection system is constructed to realize full coverage of multi-source data: at the space-based level, a UAV carrying a hyperspectral camera and a thermal infrared imager is used to plan a flight path in combination with the tidal cycle of the coastal wetland, the flight height is controlled at 100-150 m to ensure a spatial resolution of 0.5-1 m, and the fine spectral characteristics of the wetland vegetation canopy and the ground temperature data are collected synchronously to capture the correlation information of the vegetation growth and the microenvironment; at the space-based level, Sentinel-2 satellite and Gaofen-6 satellite are selected as macro data sources to batch acquire macro ecological information such as global vegetation coverage, intertidal zone landform change and water body distribution range of the monitoring area, and make up for the deficiency of limited UAV monitoring range; at the ground-based level, an Internet of Things sensor array is arranged according to the principle of uniform distribution + key area encryption, and monitoring points are arranged in different vegetation coverage areas and soil type areas of the wetland, the sensor collects soil moisture content, soil conductivity and air temperature and humidity data in real time, the sampling frequency is set to 1 time / 30 min, and the data are transmitted to the data center in real time through the NB-IoT low-power network to ensure the continuity of the soil environment parameters; at the water-based level, water quality monitoring buoys with autonomous navigation function are launched, the buoys are equipped with pH sensors, dissolved oxygen sensors, total nitrogen detectors and total phosphorus detectors, and collect key indicators of water bodies at a frequency of 1 time / 15 min, the navigation path covers the main ditches of the wetland, the intertidal zone and the near-shore shallow water area, the data are high-speed returned through the 5G network, the buoy position information is recorded synchronously to match the spatial coordinates of the water sampling points, and finally a multi-dimensional original data set covering the wetland vegetation, soil, air and water body is formed.

[0019] In step 102, a wavelet threshold denoising algorithm based on an attention mechanism is used to process satellite remote sensing data, and a space-time attention fusion model is used for data space-time alignment to obtain preprocessed data. In this embodiment, a db4 wavelet basis is selected as a decomposition basis function to perform multi-scale wavelet decomposition on the satellite remote sensing data to obtain low-frequency approximation coefficients and high-frequency detail coefficients at different scales; the high-frequency detail coefficients are input into an attention calculation network, the features are mapped through a fully connected layer, and then normalized by a Softmax function to obtain the attention weight corresponding to each high-frequency detail coefficient; the high-frequency detail coefficients are processed based on the attention weight, the processed high-frequency detail coefficients and the original low-frequency approximation coefficients are reconstructed by wavelet to obtain the denoised satellite remote sensing data.

[0020] In this embodiment, the coastal wetland monitoring area is divided into a plurality of spatial grids in the spatial dimension and a plurality of 1-hour time segments in the time dimension, forming a grid system covering the whole space-time of the monitoring area; the processed data of the unmanned aerial vehicle, satellite remote sensing data, Internet of Things sensor data, and water quality monitoring buoy data are subjected to space-time feature extraction; the space-time features are input into the space-time attention calculation module, classified by time segment, and in each time segment, the correlation between the space-time features and the grid system is calculated, and according to the correlation size, the attention weight of the space-time features in the corresponding grid is calculated through the full connection layer and the Softmax function; the attention weights are weighted and summed to obtain the comprehensive data of the space-time grid in the corresponding time segment, and all space-time grids and time segments are traversed to form the preprocessed data.

[0021] Step 103, input the preprocessed data into the Transformer-ResNet hybrid model for environmental change evaluation, and output the ecological health index, wherein the Transformer-ResNet hybrid model includes an adaptive multi-scale space-time attention module, a multi-task learning branch, and a dynamic weight adjustment mechanism; In this embodiment, the preprocessed data after space-time alignment is input into the Transformer-ResNet hybrid model, the spatial features of the data are extracted through the convolution layer of the ResNet network, and the data is converted into a feature vector conforming to the processing format of the Transformer through the embedding layer of the Transformer network; the ResNet network adopts a 3x3 size convolution kernel to perform multi-round convolution operation on the data, and after each round of convolution, the feature distribution is stabilized through batch normalization processing, and the gradient disappearance problem of the deep network is relieved through residual connection, focusing on extracting local spatial features in the data, such as texture details of wetland vegetation canopy, boundary contours of soil and water, spatial distribution patterns of small landforms, etc. At the same time, the embedding layer of the Transformer network first converts the structured data after space-time alignment, such as the environmental parameter values of each grid, into a high-dimensional vector, and then adds position encoding containing spatial coordinates and timestamp information, so that the vector carries both the numerical features and the space-time position features of the data, forming a sequence feature vector suitable for Transformer processing; The spatial features extracted by ResNet and the feature vectors embedded by Transformer are input into the adaptive multi-scale spatio-temporal attention module. Through multi-scale feature extraction and adaptive weight allocation mechanism, the spatio-temporal features of different scales are automatically learned to obtain multi-scale spatial attention weights. The multi-scale feature extraction part sets three sliding windows of 1 day, 3 days and 7 days for the time dimension, respectively capturing short-term features such as daily tidal changes, medium-term features such as weekly vegetation growth fluctuations, and long-term features such as half-month scale water quality trend time features. For the spatial dimension, three scale grids of 50m×50m, 100m×100m and 200m×200m are divided to extract local features such as single vegetation patch, regional features such as vegetation community, and global features such as the whole wetland spatial features. The adaptive weight allocation mechanism assigns dynamic weights to each scale feature by calculating the relevance of different scale features to wetland ecological changes, such as the higher relevance of 3-day scale vegetation features to ecological health in the growing season. The higher the relevance, the greater the feature weight. Finally, multi-scale spatial attention weights are formed by integrating them to strengthen the representation of key spatio-temporal features. The features adjusted by attention are input into the multi-task learning branch. Each branch uses a fully connected layer to map the features to obtain the numerical values of the four environmental evaluation indicators. The multi-task learning branch includes vegetation coverage evaluation branch, water quality compliance rate evaluation branch, soil salinization degree evaluation branch, and hydrological connectivity evaluation branch. Specifically, the vegetation coverage evaluation branch maps the features to the [0, 1] interval through a fully connected layer, calculates the vegetation pixel ratio based on the vegetation index such as NDVI in the unmanned aerial vehicle hyperspectral data, and outputs the vegetation coverage value. The water quality compliance rate evaluation branch refers to the surface water quality standard such as TP≤0.2mg / L for Class III water based on the feature mapping result, and calculates the proportion of compliant sampling points to output the water quality compliance rate. The soil salinization degree evaluation branch maps the features to a quantitative score of 0-10, where 0 represents no salinization and 10 represents severe salinization, and combines the soil conductivity data to divide the grades and convert them into numerical values. The hydrological connectivity evaluation branch calculates the connectivity probability of each water node in the wetland by feature mapping and outputs a connectivity index in the range of 0-1. Each branch shares the underlying features during the training process and is optimized simultaneously through backpropagation to improve the evaluation accuracy of individual indicators. Based on the wetland type, seasonal changes and historical ecological data, the weights of each environmental assessment index are automatically adjusted, and the prediction results of the multi-task learning branch are fused by weighting to generate the final ecological health index. First, the system calls the preset basic weight according to the wetland type, such as estuary wetland, and gives the soil salinization index a higher basic weight in the tidal flat wetland. Then, combined with seasonal information for dynamic adjustment, such as increasing the weights of hydrological connectivity and water quality compliance rate in the rainy season, and increasing the weight of vegetation coverage in the growing season. At the same time, historical ecological data such as the same period ecological health key driving factor in the past 3 years are introduced, and the weights are corrected by the attention calculation unit to make the weight distribution more in line with the actual ecological law of the wetland, such as historical data showing that the vegetation in a certain area is the core driving factor of ecological health, which further increases its weight. Finally, the numerical values of the four environmental assessment indexes are weighted and summed according to the adjusted weights, and the weight sum is 1, to obtain the ecological health index in the range of 0-1, 1 indicates that the ecological state is excellent, and 0 indicates that it is poor, realizing the comprehensive quantitative evaluation of the ecological health state of the wetland.

[0022] The adaptive weight distribution layer includes an attention mechanism submodule. The submodule is composed of a fully connected layer and an activation function, which is used to learn the weights of different scale features. The input is each scale feature output by the multi-scale feature extraction layer, which is mapped through the fully connected layer and normalized to weight values through the activation function (such as Softmax). These weight values reflect the importance of different scale features in the overall feature.

[0023] The feature fusion layer fuses each scale feature after weight distribution. The weighted sum is used to multiply each scale feature by its corresponding weight value and add it to obtain the fused feature representation, which integrates the spatial and temporal information of different scales.

[0024] Feature extraction implementation: In terms of spatial feature extraction, each convolution branch of the multi-scale feature extraction layer performs convolution operation on the input data. Taking the 3x3 convolution kernel branch as an example, it extracts features in a small local area and can capture detailed information; while the 7x7 convolution kernel branch covers a larger area and extracts more global spatial features. In terms of time feature extraction, one-dimensional convolution operation slides along the time dimension, and different lengths of time window can capture short-term and long-term time series change patterns.

[0025] Weight distribution implementation: The adaptive weight distribution layer receives the features output by the multi-scale feature extraction layer. First, the features of each scale are flattened and input into a fully connected layer, which performs a preliminary transformation on the features by learning the complex nonlinear relationship between the features and the weights. Then, after processing by an activation function, the output is limited to an appropriate range and normalized as weights. For example, the Softmax function can ensure that each weight value is between 0 and 1, and the sum of all weights is 1, so that reasonable weights are automatically assigned according to the contribution of different scale features to the overall task. Finally, the features of each scale are fused into a more representative feature representation through a weighted sum operation of the feature fusion layer, which is used for subsequent environmental change evaluation.

[0026] Step 104, when the predicted ecological health index is lower than the threshold value, the PPO algorithm is used to dynamically adjust the monitoring strategy according to the warning level and push the warning report.

[0027] In this embodiment, when the predicted ecological health index is lower than the threshold value, the current warning level and the corresponding ecological health index value are recorded as the initial input information of the PPO algorithm. If the ecological health index is between 0.5 and 0.6, it is determined as a yellow warning; if the ecological health index is between 0.4 and 0.5, it is determined as an orange warning; and if the ecological health index is lower than 0.4, it is determined as a red warning. The policy network and value network of the PPO algorithm are initialized, and a plurality of candidate monitoring strategies are generated according to the current warning level and the initialized policy network. The policy network uses a 3-layer fully connected neural network structure, the input layer includes the current warning level represented in one-hot encoding form, the existing monitoring strategy parameters such as the unmanned aerial vehicle inspection frequency, the sensor sampling interval, and the number of buoy monitoring points, and the output layer is the specific value of the adjusted strategy parameters such as the frequency, interval, and point. The value network is also a 3-layer fully connected network, the input layer includes the warning level and the current wetland environment state such as the ecological health index and the main abnormal factors, and the output layer is the estimated value of the strategy execution effect. During initialization, the policy network parameters are randomly assigned, and the value network parameters are pre-trained based on historical strategy evaluation data. According to the current warning level such as yellow, orange, and red, the policy network generates candidate strategies within the preset parameter range: for example, when the yellow warning is generated, the unmanned aerial vehicle inspection frequency is adjusted from 6 hours / once to 4 hours / once, 3 hours / once, etc. 5 candidate schemes; when the red warning is generated, 8 candidate schemes including the buoy sampling frequency is increased to 5 minutes / once, the unmanned aerial vehicle inspection frequency in the key area is doubled, etc. are generated to ensure that different intensity of strategy adjustment directions are covered. The candidate strategies are applied to the simulation monitoring scene, the environmental state changes during the execution of the strategies are recorded, and the performance scores of each group of strategies are evaluated by the value network; the simulation monitoring scene is constructed based on historical monitoring data of wetlands, and contains typical environmental characteristics under different warning levels, such as light degradation of vegetation data under yellow warning and serious over-standard water quality data under red warning, and the data flow changes after the execution of the strategies can be reproduced; when the candidate strategies are applied, the system simulates the process of unmanned aerial vehicle inspection at a new frequency and sensor sampling at a new interval, and records key environmental state indicators, such as whether the spatio-temporal coverage of data collection covers all abnormal areas, the abnormal information capture delay from the occurrence of the abnormality to the monitoring, and the data redundancy rate of invalid repeated data; the value network outputs the performance score of each group of candidate strategies based on these state indicators and the benchmark value of the historical optimal strategy, and the score is 0-100, and the higher the score is, the more reasonable the resource consumption is, such as high coverage, low delay and low redundancy, while ensuring the monitoring accuracy; According to the performance score of the candidate strategy, the advantage function of the strategy is calculated, and the parameter of the strategy network is updated through the clipping target function of the PPO algorithm; the advantage function calculation formula is the difference between the performance score of the current candidate strategy and the average score of all candidate strategies, and a positive value indicates that the strategy is better than the average level, and the larger the difference is, the more significant the advantage is; the clipping target function is used to limit the updating range of the strategy, so as to avoid the violent fluctuation of the network parameters caused by the single evaluation error: first, the probability ratio of the new strategy and the old strategy is calculated to reflect the change range of the strategy, then the ratio is limited in the interval [1-ε, 1+ε], and ε is usually set to 0.2, and finally the loss value is calculated by combining the advantage function; when the advantage is positive, the smaller upper limit of the ratio is used to calculate the loss; when the advantage is negative, the larger lower limit of the ratio is used to calculate the loss; the loss value is minimized through the gradient descent algorithm, and the weight and bias parameters of the strategy network are updated, so that the network tends to generate strategies with significant advantages; After multiple iterations, the optimal monitoring strategy under the current warning level is determined, and a warning report is generated and pushed according to the finally determined warning level and the optimal monitoring strategy; during the iteration process, the candidate strategy-simulation evaluation-update network process is repeated every round until the performance score of the optimal strategy in the last three consecutive iterations fluctuates less than 5%, that is, the strategy performance tends to be stable; at this time, the candidate strategy with the highest score in the last iteration is selected as the optimal monitoring strategy, which needs to meet the following conditions at the same time: the monitoring coverage rate of the abnormal area is greater than or equal to 95%, the abnormal capture delay is less than or equal to 30 minutes, and the data redundancy rate is less than or equal to 10%; based on the optimal strategy and the warning level, the system automatically generates a warning report, including: the warning level and the triggering reason such as the ecological health index 0.45 triggering an orange warning, the main reason for the TN concentration exceeding the standard, the details of the optimal monitoring strategy such as the sampling frequency of the buoy being adjusted to 5 minutes / time, the key inspection area of the unmanned aerial vehicle being the eastern intertidal zone, and the recommended remediation measures such as throwing water purifying agents and limiting the discharge of surrounding pollution sources; the report is pushed to the coastal wetland management platform in real time through the API interface, and is sent to the management personnel terminal at the same time, ensuring fast response.

[0028] Please refer to Figure 2 The structure diagram of the intelligent monitoring system for coastal wetlands based on artificial intelligence provided by the embodiment of the present application, the system comprises: The acquisition module is configured to acquire unmanned aerial vehicle data, satellite remote sensing data, Internet of Things sensor data, and water quality monitoring buoy data of the coastal wetland. The preprocessing module is configured to process the satellite remote sensing data by using a wavelet threshold denoising algorithm based on an attention mechanism, and perform data space-time alignment based on a space-time attention fusion model to obtain preprocessed data. The evaluation module is configured to input the preprocessed data into a Transformer-ResNet hybrid model to evaluate environmental changes and output an ecological health index, wherein the Transformer-ResNet hybrid model comprises an adaptive multi-scale space-time attention module, a multi-task learning branch, and a dynamic weight adjustment mechanism. The monitoring module is configured to dynamically adjust a monitoring strategy according to a warning level by using a PPO algorithm and push a warning report when the predicted ecological health index is lower than a threshold.

[0029] Figure 3is a structural schematic view of an intelligent monitoring device for coastal wetlands based on artificial intelligence provided by an embodiment of the present application. The intelligent monitoring device for coastal wetlands based on artificial intelligence 300 can have great differences due to different configurations or performances, and can include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, one or more storage media 330 (for example, one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and the storage medium 330 can be temporary storage or persistent storage. The programs stored in the storage medium 330 can include one or more modules (not shown in the figure), and each module can include a series of instruction operations in the intelligent monitoring device for coastal wetlands based on artificial intelligence 300. Further, the processor 310 can be configured to communicate with the storage medium 330 and execute a series of instruction operations in the storage medium 330 on the intelligent monitoring device for coastal wetlands based on artificial intelligence 300 to implement the method provided by the above embodiment.

[0030] The intelligent monitoring device for coastal wetlands based on artificial intelligence 300 can also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art can understand that, Figure 3 The structure of the intelligent monitoring device for coastal wetlands based on artificial intelligence shown does not constitute a limitation on the computer device provided by the present application, and can include more or fewer components than shown, or combine certain components, or different component arrangements.

[0031] The present application also provides a computer readable storage medium, which can be a non-volatile computer readable storage medium or a volatile computer readable storage medium. The computer readable storage medium has instructions stored therein, and when the instructions are run on a computer, the computer executes the steps of the intelligent monitoring method for coastal wetlands based on artificial intelligence provided by the above embodiments.

[0032] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described device or apparatus, unit can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0033] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0034] The basic principles, main features and advantages of the present application are shown and described above. Those skilled in the art should understand that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based intelligent monitoring method for coastal wetlands, characterized in that, The method comprises the following steps: Collecting unmanned aerial vehicle data, satellite remote sensing data, Internet of Things sensor data and water quality monitoring buoy data of the coastal wetland; Satellite remote sensing data is processed by a wavelet threshold denoising algorithm based on an attention mechanism, and data spatio-temporal alignment is performed based on a spatio-temporal attention fusion model to obtain preprocessed data; The preprocessed data is input into a Transformer-ResNet hybrid model for environmental change evaluation, and an ecological health index is output, wherein the Transformer-ResNet hybrid model comprises an adaptive multi-scale spatio-temporal attention module, a multi-task learning branch and a dynamic weight adjustment mechanism; When the predicted ecological health index is lower than a threshold value, a PPO algorithm is used to dynamically adjust the monitoring strategy according to the warning level and push a warning report; The preprocessed data after spatio-temporal alignment is input into the Transformer-ResNet hybrid model, and through a multi-scale feature extraction and adaptive weight allocation mechanism, spatio-temporal features of different scales are automatically learned to obtain multi-scale spatial attention weights; The features adjusted by attention are input into the multi-task learning branch to obtain numerical values of four environmental evaluation indexes; The final ecological health index is generated by weightedly fusing the prediction results of the multi-task learning branch. 2.The AI-based intelligent monitoring method for coastal wetlands according to claim 1, wherein, The satellite remote sensing data is processed by a wavelet threshold denoising algorithm based on an attention mechanism, and data spatio-temporal alignment is performed based on a spatio-temporal attention fusion model to obtain preprocessed data; A db4 wavelet basis is selected as a decomposition basis function to perform multi-scale wavelet decomposition on the satellite remote sensing data to obtain low-frequency approximation coefficients and high-frequency detail coefficients at different scales; The high-frequency detail coefficients are input into an attention calculation network, the features are mapped through a fully connected layer, and then normalized by a Softmax function to obtain attention weights corresponding to each high-frequency detail coefficient; The high-frequency detail coefficients are thresholded based on the attention weights, the processed high-frequency detail coefficients and the original low-frequency approximation coefficients are reconstructed by wavelets to obtain denoised satellite remote sensing data. 3.The AI-based intelligent monitoring method for coastal wetlands according to claim 1, wherein, The data spatio-temporal alignment based on the spatio-temporal attention fusion model to obtain the preprocessed data comprises: The monitoring area of the coastal wetland is divided into a plurality of spatial grids in the spatial dimension and a plurality of 1-hour time segments in the time dimension to form a grid system covering the whole space-time of the monitoring area; Spatio-temporal features are extracted from the processed unmanned aerial vehicle data, satellite remote sensing data, Internet of Things sensor data and water quality monitoring buoy data; The spatio-temporal features are input into a spatio-temporal attention calculation module, classified by time segments, and in each time segment, the correlation between the spatio-temporal features and the grid system is calculated, and the attention weights of the spatio-temporal features in the corresponding grid are calculated by a fully connected layer and a Softmax function according to the correlation size; The attention weights are weightedly summed to obtain the comprehensive data of the spatio-temporal grid in the corresponding time segment, and all spatio-temporal grids and time segments are traversed to form the preprocessed data. 4.The method of claim 1, wherein the method comprises: The multi-task learning branch comprises a vegetation coverage evaluation branch, a water quality compliance rate evaluation branch, a soil salinization degree evaluation branch and a hydrological connectivity evaluation branch. 5.The method of claim 1, wherein the method comprises: The PPO algorithm is adopted to dynamically adjust the monitoring strategy according to the warning level and push a warning report when the predicted ecological health index is lower than a threshold value, including: When the predicted ecological health index is lower than the threshold value, the current warning level and the corresponding ecological health index value are recorded as the initial input information of the PPO algorithm; The strategy network and the value network of the PPO algorithm are initialized, and a plurality of groups of candidate monitoring strategies are generated according to the current warning level and the initialized strategy network; Each group of candidate strategies is applied to a simulated monitoring scene, the environmental state changes during the execution of the strategy are recorded, and the performance score of each group of strategies is evaluated by the value network; The advantage function of the strategy is calculated according to the performance score of the candidate strategy, and the parameter of the strategy network is updated by the clipping objective function of the PPO algorithm; After a plurality of iterations, the optimal monitoring strategy under the current warning level is determined, and a warning report is generated and pushed according to the finally determined warning level and the optimal monitoring strategy.

6. The artificial intelligence-based intelligent monitoring method for coastal wetlands according to claim 5, characterized in that, If the ecological health index is between 0.5 and 0.6, it is determined as yellow warning; if the ecological health index is between 0.4 and 0.5, it is determined as orange warning; If the ecological health index is lower than 0.4, it is determined as red warning.

7. The intelligent monitoring system for coastal wetlands based on artificial intelligence, characterized in that, The system comprises: A collection module for collecting unmanned aerial vehicle data, satellite remote sensing data, Internet of Things sensor data and water quality monitoring buoy data of the coastal wetland; A preprocessing module for processing satellite remote sensing data by using a wavelet threshold denoising algorithm based on an attention mechanism, performing data space-time alignment based on a space-time attention fusion model, and obtaining preprocessed data; an evaluation module for inputting the preprocessed data into a Transformer-ResNet hybrid model for environmental change evaluation, and outputting an ecological health index, wherein the Transformer-ResNet hybrid model comprises an adaptive multi-scale space-time attention module, a multi-task learning branch and a dynamic weight adjustment mechanism; the preprocessed data after space-time alignment is input into the Transformer-ResNet hybrid model, the multi-scale feature extraction and adaptive weight distribution mechanism are used to automatically learn the space-time features of different scales, and the multi-scale spatial attention weight is obtained; the features adjusted by attention are input into the multi-task learning branch to obtain the values of four environmental evaluation indexes; the final ecological health index is generated by weightedly fusing the prediction results of the multi-task learning branch; A monitoring module for dynamically adjusting the monitoring strategy according to the warning level by using the PPO algorithm and pushing a warning report when the predicted ecological health index is lower than a threshold value.

8. The intelligent monitoring equipment for coastal wetlands based on artificial intelligence, characterized in that, The artificial intelligence-based coastal wetland intelligent monitoring device comprises a memory and at least one processor, and the memory stores instructions; the at least one processor calls the instructions in the memory to enable the artificial intelligence-based coastal wetland intelligent monitoring device to perform each step of the artificial intelligence-based coastal wetland intelligent monitoring method according to any one of claims 1-6.

9. A computer-readable storage medium having stored thereon instructions, the computer-readable storage medium comprising: The instructions are executed by the processor to implement each step of the artificial intelligence-based coastal wetland intelligent monitoring method according to any one of claims 1-6.

Images