Intelligent monitoring system for coastal wetland ecological parameters based on deep learning

Abstract

The application discloses a coastal wetland ecological parameter intelligent monitoring system based on deep learning, and relates to the technical field of wetland ecological parameter monitoring.The system comprises a multi-source monitoring data access and preprocessing module, a coastal wetland ecological mechanism knowledge base construction module, a knowledge-data double-driven deep learning inversion module, a multi-type wetland self-adaptive adaptation module and an ecological parameter visualization and output module.The application converts the ecological process mechanism of vegetation-soil-hydrology coupling into a differentiable constraint term embedded in a deep learning network by constructing a coastal wetland ecological mechanism knowledge base, realizes deep fusion of ecological mechanism prior knowledge and data-driven models, effectively guides the feature learning direction of the model, avoids abnormal results that do not conform to ecological laws output by pure data-driven models, and improves the model interpretability and the physical rationality of the inversion results.The application further constructs a multi-branch self-adaptive feature extraction network that is suitable for different wetland types.

Description

Technical Field

[0001] This invention relates to the field of wetland ecological parameter monitoring technology, specifically to an intelligent monitoring system for coastal wetland ecological parameters based on deep learning. Background Technology

[0002] Coastal wetlands are unique ecosystems located in the transitional zone between land and sea, encompassing various types such as intertidal mudflats, salt marshes, mangrove wetlands, and estuarine wetlands. They are among the most productive and biodiverse ecosystems on Earth, exhibiting characteristics of both terrestrial and marine ecosystems. They play irreplaceable ecological roles, including wind and wave protection, water purification, biodiversity maintenance, and carbon sequestration. Coastal wetland ecological parameter monitoring involves quantitative, routine, and spatiotemporally continuous monitoring and inversion of the three core elements of coastal wetland ecosystems: vegetation, soil, and hydrology. This is a core technological means to comprehensively understand the structure, function, and spatiotemporal evolution of coastal wetland ecosystems. Accurate, efficient, and large-scale monitoring of coastal wetland ecological parameters can promptly identify degradation trends, human disturbances, and potential ecological risks, scientifically assess the effectiveness of wetland protection and restoration projects, and provide scientific support for the formulation and optimization of wetland protection and management policies.

[0003] Deep learning technology, with its powerful capabilities in feature extraction from multi-source heterogeneous data, fitting of nonlinear relationships, and spatiotemporal sequence modeling, can effectively process various types of information acquired during coastal wetland monitoring, such as remote sensing images, UAV aerial survey data, and in-situ station monitoring data. It breaks through the bottlenecks of traditional methods such as manual interpretation and empirical model inversion, which suffer from low efficiency, limited coverage, insufficient inversion accuracy, and difficulty in achieving continuous spatiotemporal monitoring. Deep learning provides a new technical path for large-scale, high spatiotemporal resolution, and automated intelligent monitoring of coastal wetland ecological parameters, and has become the core technology development direction in the current field of coastal wetland ecological monitoring.

[0004] However, existing deep learning-based coastal wetland ecological parameter monitoring technologies mostly adopt a purely data-driven network model architecture, relying solely on labeled sample data to complete model training and parameter inversion. They fail to incorporate constraints based on the ecological processes of coastal wetlands themselves. This makes the model's feature learning direction susceptible to interference from data noise and sample bias, easily leading to abnormal inversion results that do not conform to the coupled ecological laws of "vegetation-soil-hydrology" in coastal wetlands. This results in insufficient interpretability of the model and insufficient physical rationality of the inversion results. Furthermore, existing models often use a uniform network structure for feature extraction and parameter inversion, failing to differentiate and adapt to the ecological characteristics of different wetland types such as mangroves, salt marshes, and intertidal mudflats. In wetland areas with scarce samples and high environmental heterogeneity, the model's generalization ability and transfer adaptability are poor, failing to meet the practical application needs of large-scale, multi-type coastal wetland routine, high-precision intelligent monitoring. Therefore, developing a deep learning-based intelligent monitoring system for coastal wetland ecological parameters is of great significance. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an intelligent monitoring system for coastal wetland ecological parameters based on deep learning. It can solve the technical problems of poor interpretability, insufficient physical rationality of inversion results, poor adaptability of existing pure data-driven coastal wetland monitoring models, weak generalization ability of models to different types of wetlands and scarce sample areas, which cannot meet the technical needs of high-precision routine monitoring of large-scale and multi-type coastal wetlands.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a deep learning-based intelligent monitoring system for coastal wetland ecological parameters. The system includes: a multi-source monitoring data access and preprocessing module, a coastal wetland ecological mechanism knowledge base construction module, a knowledge-data dual-driven deep learning inversion module, a multi-type wetland adaptive adaptation module, and an ecological parameter visualization and output module. The modules interact with each other through a standardized data interface.

[0007] The multi-source monitoring data access and preprocessing module accesses multi-source heterogeneous monitoring data of coastal wetlands, and after completing standardized preprocessing, generates a standardized monitoring dataset, which is then output to the coastal wetland ecological mechanism knowledge base construction module and the knowledge-data dual-driven deep learning inversion module, respectively.

[0008] The coastal wetland ecological mechanism knowledge base construction module is based on the coastal wetland vegetation-soil-hydrology coupled ecological process mechanism to construct a prior constraint knowledge base, transforming the ecological mechanism rules into differentiable mechanism constraint terms, and outputting them to the knowledge-data dual-driven deep learning inversion module and the multi-type wetland adaptive adaptation module respectively.

[0009] The knowledge-data dual-driven deep learning inversion module receives a standardized monitoring dataset and mechanism constraint terms, completes feature extraction, mechanism constraint embedding and end-to-end inversion of ecological parameters, and outputs the initial inversion results to the multi-type wetland adaptive adaptation module.

[0010] The multi-type wetland adaptive adaptation module has a built-in multi-branch adaptive feature extraction network to complete the automatic classification of wetland types and the optimization of differential parameters, and outputs high-precision ecological parameter inversion results to the ecological parameter visualization and output module.

[0011] The ecological parameter visualization and output module completes the visualization display, accuracy assessment and data export of the inversion results.

[0012] Furthermore, the multi-source monitoring data access and preprocessing module performs the following operations during data access and preprocessing:

[0013] It integrates satellite multispectral remote sensing image data, satellite hyperspectral remote sensing image data, UAV aerial survey image data, UAV LiDAR point cloud data, in-situ hydrological monitoring data, in-situ soil monitoring data, in-situ vegetation monitoring data, and tidal hydrological observation data, and completes the format unification and storage of various types of data.

[0014] For optical image data, radiometric calibration, atmospheric correction, geometric fine correction, and noise removal are performed; for in-situ site data, outlier removal, missing value interpolation, and time-series alignment are performed.

[0015] Spatial registration and data normalization are performed on all preprocessed data to generate a standardized monitoring dataset with matching spatiotemporal dimensions.

[0016] Furthermore, the prior constraint knowledge base constructed by the coastal wetland ecological mechanism knowledge base construction module includes the mapping relationship rules between vegetation spectral response characteristics and physiological and ecological parameters corresponding to different coastal wetland types, the mapping relationship rules between vegetation spectral response characteristics and environmental driving factors, the ecological constraint thresholds of core ecological parameters under different tidal conditions, and the spatiotemporal variation patterns of core ecological parameters under different seasonal cycles. The module transforms all ecological mechanism rules in the prior constraint knowledge base into differentiable mechanism constraint terms that can be embedded in deep learning networks. The differentiable mechanism constraint terms are divided into two categories: feature layer constraint terms and loss layer constraint terms.

[0017] Furthermore, the coastal wetland ecological mechanism knowledge base construction module performs the following operations when constructing the prior constraint knowledge base and mechanism constraint items:

[0018] By integrating the mechanistic research results of the coupled ecological processes of vegetation-soil-hydrology in coastal wetlands with long-term in-situ observation data, the mapping relationship rules between vegetation spectral response characteristics and physiological and ecological parameters and environmental driving factors are extracted.

[0019] Based on the differences in ecological characteristics of different types of coastal wetlands, corresponding ecological units are divided, the ecological constraint thresholds and spatiotemporal variation patterns of core ecological parameters within each ecological unit are determined, and a priori constraint knowledge base is constructed.

[0020] The mapping rules and constraint thresholds in the prior constraint knowledge base are transformed into differentiable mathematical expressions to generate feature layer constraint terms and loss layer constraint terms. The loss layer constraint terms are calculated using the following formula: ,in: The ecological mechanism constraint loss value is N, where N is the total number of ecological parameter types. The mechanistic constraint weights for the i-th type of ecological parameter are determined based on the statistical characteristics of long-term ecological observation data of coastal wetlands. The values ​​of the i-th type of ecological parameters retrieved from the model are as follows: The ecological mechanism threshold of the i-th type of ecological parameter in the prior constraint knowledge base;

[0021] The generated two types of mechanism constraint terms are format-standardized and output to the knowledge-data dual-driven deep learning inversion module.

[0022] Furthermore, the knowledge-data dual-driven deep learning inversion module includes a feature extraction submodule, a mechanism constraint embedding submodule, and an end-to-end parameter inversion submodule. The feature extraction submodule adopts a multi-level convolutional neural network architecture. The mechanism constraint embedding submodule is connected to the coastal wetland ecological mechanism knowledge base construction module through a standardized data interface. The end-to-end parameter inversion submodule adopts a multi-branch decoding structure, and the number of branches in the multi-branch decoding structure is consistent with the number of ecological parameter types to be inverted.

[0023] Furthermore, the knowledge-data dual-driven deep learning inversion module performs the following operations during feature extraction and parameter inversion:

[0024] The feature extraction submodule receives a standardized monitoring dataset and extracts spectral features, texture features, and spatiotemporal correlation features of the input data through a multi-level convolutional neural network to generate a multi-level initial feature map.

[0025] The mechanism constraint embedding submodule receives differentiable mechanism constraint terms, embeds the feature layer constraint terms into the multi-level feature extraction stage of the feature extraction submodule, performs constraint optimization on the multi-level initial feature map, and generates the optimized feature map;

[0026] The mechanism constraint embedding submodule weights and fuses the loss layer constraint terms with the data fitting loss terms to construct the joint loss function of the model. The joint loss function is calculated using the following formula: ,in, This represents the total loss value of the model. Fit loss values ​​to the model data. The ecological mechanism constraint loss value is α, which is the loss fusion coefficient. This coefficient is determined through adaptive iterative optimization of the ecological parameter inversion accuracy during model training.

[0027] The end-to-end parameter inversion submodule receives the optimized feature map, completes feature decoding and spatial dimension restoration through a multi-branch decoding structure, and outputs the initial inversion results corresponding to various ecological parameters.

[0028] Furthermore, the multi-branch adaptive feature extraction network built into the multi-type wetland adaptive adaptation module includes a first branch network corresponding to mangrove wetlands, a second branch network corresponding to salt marsh wetlands, and a third branch network corresponding to intertidal mudflats. Each branch network adopts independent convolutional kernel size, network layer number, and feature weight parameters. The structural parameters of each branch network are matched with the ecological features of the corresponding wetland type, and each branch network is equipped with an independent transfer learning adaptation interface.

[0029] Furthermore, the multi-type wetland adaptive adaptation module performs the following operations when classifying wetland types and optimizing parameters:

[0030] Receive standardized monitoring datasets and initial inversion results, and based on the wetland type classification rules in the prior constraint knowledge base, complete the automatic identification and spatial boundary delineation of mangrove wetlands, salt marsh wetlands, and intertidal mudflats within the monitoring area, and generate regional masks corresponding to each wetland type;

[0031] Based on the regional mask of each wetland type, the optimized feature map of the corresponding region and the initial inversion result are input into the matching branch network to complete the secondary extraction of the corresponding region features and the correction of the inversion result;

[0032] For wetland areas where sample coverage is below a preset threshold, lightweight adaptation and weight fine-tuning of the corresponding branch network are completed through the transfer learning adaptation interface. The weight fine-tuning is calculated using the following formula: ,in, To fine-tune the branch network weights, For the pre-trained weights of the source wetland type, β represents the initial random weights for the target wetland area, and β is the weight fusion ratio, which is calculated and determined based on the sample coverage and ecological feature similarity of the target area.

[0033] The inversion results output from each branch network are spliced ​​and merged according to the spatial boundaries to generate high-precision ecological parameter inversion results for the entire monitoring area.

[0034] Furthermore, the ecological parameter visualization and output module incorporates a thematic map generation unit, a time series analysis unit, an accuracy assessment unit, and a data export unit. The thematic map generation unit receives high-precision ecological parameter inversion results and generates spatial distribution raster maps and vector thematic maps corresponding to various ecological parameters. The time series analysis unit performs time series variation statistics and trend fitting of multiple inversion results. The accuracy assessment unit performs error statistics and accuracy verification of the inversion results based on in-situ measured data. The data export unit supports exporting inversion data in various standardized formats and calling external interfaces.

[0035] Furthermore, the output of the multi-source monitoring data access and preprocessing module is connected to the input of the coastal wetland ecological mechanism knowledge base construction module and the input of the knowledge-data dual-driven deep learning inversion module. The output of the coastal wetland ecological mechanism knowledge base construction module is connected to the input of the knowledge-data dual-driven deep learning inversion module and the input of the multi-type wetland adaptive adaptation module. The output of the knowledge-data dual-driven deep learning inversion module is connected to the input of the multi-type wetland adaptive adaptation module. The output of the multi-type wetland adaptive adaptation module is connected to the input of the ecological parameter visualization and output module. The standardized data interfaces between the modules adopt a unified communication protocol and data format specification.

[0036] Compared with existing technologies, this deep learning-based intelligent monitoring system for coastal wetland ecological parameters has the following advantages:

[0037] This invention constructs a knowledge base of coastal wetland ecological mechanisms, transforming the ecological process mechanism of vegetation-soil-hydrology coupling into differentiable constraint terms embedded in a deep learning network. This achieves a deep integration of prior knowledge of ecological mechanisms with a data-driven model, effectively guiding the model's feature learning direction and avoiding abnormal results that do not conform to ecological laws from purely data-driven models. This improves the interpretability of the model and the physical rationality of the inversion results. By constructing a multi-branch adaptive feature extraction network adapted to different wetland types and combining it with a transfer learning strategy based on mechanism constraints, the model achieves differentiated adaptation to multiple types of coastal wetlands. This significantly improves the model's generalization ability and transfer adaptability in areas with scarce samples, meeting the practical application needs of high-precision, routine, and intelligent monitoring of large-scale coastal wetlands.

[0038] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0040] Figure 1 This is a schematic diagram of the structure of a deep learning-based intelligent monitoring system for coastal wetland ecological parameters.

[0041] Figure 2 This is a flowchart of the workflow for a deep learning-based intelligent monitoring system for coastal wetland ecological parameters.

[0042] Figure 3 This is a flowchart of the multi-source monitoring data access and preprocessing module during data access and preprocessing. Detailed Implementation

[0043] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0044] The intelligent monitoring system for coastal wetland ecological parameters provided by this invention takes the coupled ecological mechanism of coastal wetland vegetation, soil, and hydrology as its core basis, integrates multi-source heterogeneous monitoring data from satellite remote sensing, UAV aerial surveys, in-situ stations, and tidal observations, completes data spatiotemporal matching through standardized preprocessing, constructs a priori constraint knowledge base based on ecological mechanisms and transforms it into differentiable constraint terms that can be embedded in deep learning networks, builds a knowledge- and data-driven inversion model to achieve initial inversion of ecological parameters, and then adapts it to the intertidal zone of mangrove salt marshes through a multi-branch adaptive feature extraction network. This system automatically classifies and optimizes the inversion results for different wetland types, including mudflats. Finally, a visualization module enables spatial display, temporal analysis, accuracy verification, and data export of ecological parameters. The entire system achieves a deep integration of prior knowledge of ecological mechanisms and a data-driven deep learning model. It solves the technical challenges of poor interpretability, insufficient physical rationality of inversion results, and weak generalization ability of traditional pure data-driven models to multi-type wetlands and areas with scarce samples. This system can meet the practical application needs of high-precision, routine, and intelligent monitoring of ecological parameters in large-scale coastal wetlands. A detailed description is provided below with reference to specific embodiments.

[0045] This embodiment selects a national nature reserve of coastal wetlands as the actual application scenario. The reserve covers a total area of ​​approximately 200 km² and features a complete distribution of three typical coastal wetland types: mangrove wetlands, salt marshes, and intertidal mudflats. It is a core area for the research and protection of coastal wetland ecosystems. This implementation relies on the system of this invention to conduct intelligent monitoring and inversion of three core ecological parameters of the reserve: vegetation cover, soil salinity, and hydrological inundation depth. The entire process is strictly executed according to the system module flow. The specific implementation process is as follows.

[0046] First, perform multi-source monitoring data access and preprocessing operations, such as... Figure 3 As shown. This step is completed by the multi-source monitoring data access and preprocessing module. The core objective is to access all multi-source heterogeneous monitoring data within the protected area and complete the standardization process to generate a standardized monitoring dataset with fully matched spatiotemporal dimensions. In the data access stage, the module simultaneously accesses multiple types of monitoring data. Among them, satellite remote sensing data includes Sentinel-2 multispectral remote sensing image data and Gaofen-5 hyperspectral remote sensing image data, with spatial resolutions of 10m and 30m, and temporal resolutions of 5d and 15d, respectively, covering the entire spatial range of the protected area. UAV monitoring data uses aerial survey image data and point cloud data acquired by a DJI M300 UAV equipped with an aerial survey camera and LiDAR sensor. The flight altitude is 120m, and the spatial resolution of the acquired data reaches 0.5m, covering key wetland monitoring sample areas within the protected area. In-situ monitoring data includes real-time hydrological data from 5 hydrological monitoring stations within the protected area, salinity and moisture content data from 8 soil monitoring profiles, coverage, plant height, and biomass data from 12 fixed vegetation plots, and hourly tidal level data from 3 tidal hydrological observation stations. The module first unifies and centrally stores all access data, storing remote sensing images in TIFF format, in-situ monitoring data in CSV format, and UAV point cloud data in LAS format, thus achieving format standardization for data from different sources.

[0047] In the data refinement stage, the module performs differentiated processing operations according to data type. For optical imagery data such as satellite multispectral, satellite hyperspectral, and UAV aerial survey images, four operations are performed sequentially: radiometric calibration, atmospheric correction, geometric fine correction, and noise removal. Radiometric calibration converts the image DN values ​​into apparent radiance; atmospheric correction uses the 6S model to eliminate errors caused by atmospheric scattering and absorption; geometric fine correction uses a high-precision topographic map of the protected area as a benchmark to control the image's geometric error to within one pixel; and noise removal uses a Gaussian filtering algorithm to remove salt-and-pepper noise and Gaussian noise from the image. For station-based data such as in-situ hydrological monitoring, in-situ soil monitoring, in-situ vegetation monitoring, and tidal hydrological observation, outlier removal, missing value interpolation, and time series alignment are performed sequentially. Outlier removal uses the 3σ criterion to remove monitoring values ​​that exceed the normal range; missing value interpolation uses linear interpolation to fill in missing data in the time series; and time series alignment uniformly calibrates all station data to the hourly Beijing time scale to ensure consistency in the time dimension.

[0048] After completing the differentiation processing, the module performs spatial registration and data normalization processing on all data. Using the projected coordinate system of satellite remote sensing imagery as a reference, all in-situ station data of UAV data are registered to the same spatial coordinate system. Then, the minimum-maximum normalization method is used to map all data values ​​to the 0~1 range, finally generating a standardized monitoring dataset with complete spatiotemporal matching. This dataset is simultaneously output to the coastal wetland ecological mechanism knowledge base construction module and the knowledge data dual-driven deep learning inversion module, providing a data foundation for subsequent processes.

[0049] Next, we will perform the operations of constructing a knowledge base for the ecological mechanisms of coastal wetlands and generating differentiable mechanism constraint terms, such as... Figure 1 As shown. This step is completed by the coastal wetland ecological mechanism knowledge base construction module. The core objective is to integrate the research results of the ecological mechanism of the protected area with long-term time-series observation data, construct a prior constraint knowledge base, and transform it into differentiable mechanism constraint terms that can be embedded in a deep learning network. The module first integrates the research results of the vegetation-soil-hydrology coupled ecological process mechanism of the protected area over the past 10 years with long-term time-series in-situ observation data, and systematically extracts the association rules of three types of core ecological parameters, including the mapping relationship rules between the vegetation spectral response characteristics of mangroves, salt marshes, intertidal mudflats and vegetation cover, the association rules between soil salinity and hydrological inundation depth, the constraint relationship rules of ecological parameters under different tidal conditions, and the spatiotemporal variation patterns of ecological parameters under the four seasons of spring, summer, autumn and winter.

[0050] Based on the differences in ecological characteristics among the three types of wetlands, the module divides the protected area into mangrove ecological units, salt marsh ecological units, and intertidal mudflat ecological units. Combining long-term observation data, the module statistically determines the ecological constraint thresholds for core ecological parameters within each ecological unit. Specifically, the constraint thresholds for mangrove vegetation coverage are 20%–95%, soil salinity are 2 g / kg–25 g / kg, and hydrological inundation depth are 0 m–1.5 m; for salt marsh vegetation coverage, the constraint thresholds are 10%–90%, soil salinity are 3 g / kg–30 g / kg, and hydrological inundation depth are 0 m–2 m; and for intertidal mudflat vegetation coverage, the constraint thresholds are 5%–50%, soil salinity are 5 g / kg–35 g / kg, and hydrological inundation depth are 0 m–2.5 m. These mapping rules and constraint thresholds together constitute the prior constraint knowledge base for coastal wetland ecology.

[0051] The module transforms all rules in the prior constraint knowledge base into differentiable mathematical expressions, generating two types of differentiable mechanism constraint terms: feature layer constraint terms and loss layer constraint terms. Feature layer constraint terms constrain the feature extraction process of the deep learning model, ensuring that the extracted features conform to ecological mechanism laws. Loss layer constraint terms constrain the model's loss calculation process, preventing the inversion results from exceeding the ecologically reasonable range. In the specific implementation of this embodiment, the loss layer constraint terms adopt the formula... Calculation. In the formula. The ecological mechanism constraint loss value is used to measure the deviation between the model inversion results and the ecological mechanism threshold. N is the total number of ecological parameter types. In this embodiment, three parameters are inverted: vegetation cover, soil salinity, and hydrological inundation depth. Therefore, N is set to 3. The weight is the mechanistic constraint weight for the i-th type of ecological parameter. This weight is determined based on the statistical characteristics of long-term ecological observation data of coastal wetlands. Specifically, the quantification method is to calculate the standard deviation of the historical observation values ​​of each type of ecological parameter, and then normalize the standard deviation values ​​as the mechanistic constraint weight of the corresponding parameter. The weight value ranges from 0 to 1. In this embodiment, the weight of vegetation cover is 0.8, the weight of soil salinity is 0.7, and the weight of hydrological inundation depth is 0.9. is the value of the i-th type of ecological parameter retrieved by the model, and is the value of the parameter to be verified output by the model. To determine the ecological mechanism threshold of the i-th type of ecological parameter in the prior constraint knowledge base, the median value of the ecological constraint range of this parameter is selected as the calculation benchmark. After the module completes the format standardization processing of the two types of constraint terms, it synchronously outputs them to the knowledge data dual-driven deep learning inversion module and the multi-type wetland adaptive adaptation module.

[0052] Then, perform a knowledge-data dual-driven deep learning inversion operation, such as... Figure 1As shown. This step is completed by the knowledge-data dual-driven deep learning inversion module. This module includes a feature extraction submodule, a mechanism constraint embedding submodule, and an end-to-end parameter inversion submodule. The core objective is to combine the standardized monitoring dataset with differentiable mechanism constraint terms to complete the end-to-end initial inversion of ecological parameters. The feature extraction submodule first receives the standardized monitoring dataset and uses a multi-level convolutional neural network architecture to perform feature extraction. The network has a total of 5 convolutional layers, extracting spectral features, texture features, and spatiotemporal correlation features of the data layer by layer. After each convolutional layer, a batch normalization layer and an activation layer are connected to improve the stability and effectiveness of feature extraction, and finally generate a multi-level initial feature map containing multi-scale information.

[0053] The mechanism constraint embedding submodule receives differentiable mechanism constraint terms and first embeds these constraint terms into the multi-layer convolutional stage of the feature extraction submodule. This optimizes the initial feature maps across multiple levels, eliminating invalid features that do not conform to the ecological mechanism and generating optimized feature maps. Subsequently, the submodule weighted and fused the loss layer constraint terms with the model data fitting loss term to construct the joint loss function. In the specific implementation of this embodiment, the joint loss function uses the formula... Calculation. In the formula. The total loss value of the model is the optimization objective for model training. The loss value for fitting the model data is calculated using the mean squared error algorithm to measure the degree of fit between the model inversion results and the measured data. The ecological mechanism constraint loss value is the ecological mechanism constraint loss calculated above. α is the loss fusion coefficient, which is determined through adaptive iterative optimization of the accuracy of ecological parameter inversion during model training. Specifically, the quantification method uses the determination coefficient and root mean square error of the validation set as the core optimization indicators, and adopts a grid search method to traverse the coefficient range of 0.1 to 0.9 with a step size of 0.1. Through multiple iterations of training, the value with the optimal accuracy of the validation set is selected as the final loss fusion coefficient. In this embodiment, the optimal loss fusion coefficient α is 0.6.

[0054] The end-to-end parameter inversion submodule receives the optimized feature map and performs parameter inversion using a multi-branch decoding structure. The number of branches is consistent with the number of ecological parameter types to be inverted. In this embodiment, three decoding branches are set up to correspond to three types of parameters: vegetation cover, soil salinity, and hydrological inundation depth. The decoding process completes feature decoding and spatial dimension restoration through deconvolution layers, and finally outputs the initial inversion results of ecological parameters for the entire protected area. This result is transmitted to the multi-type wetland adaptive adaptation module to provide a basis for subsequent optimization and correction.

[0055] Next, we will perform adaptive adaptation and inversion result optimization operations for multiple wetland types, such as... Figure 2As shown. This step is completed by a multi-type wetland adaptive adaptation module. The module has a built-in multi-branch adaptive feature extraction network, including a first branch network corresponding to mangrove wetlands, a second branch network corresponding to salt marsh wetlands, and a third branch network corresponding to intertidal mudflats. Each branch network uses independent convolutional kernel size, network layer number, and feature weight parameters to match the ecological characteristics of the corresponding wetland type. At the same time, an independent transfer learning adaptation interface is set. The module first receives the standardized monitoring dataset and the initial inversion results of ecological parameters. Based on the wetland type classification rules in the coastal wetland ecological mechanism knowledge base, it automatically identifies wetland types and delineates spatial boundaries across the entire protected area. Through a triple judgment of spectral features, texture features, and ecological thresholds, it accurately distinguishes the spatial range of mangrove wetlands, salt marsh wetlands, and intertidal mudflats, generating regional masks for the three types of wetlands.

[0056] Based on regional masks, the module inputs the optimized feature maps and initial inversion results of the corresponding wetland regions into matching branch networks. Data from mangrove areas is input into the first branch network, data from salt marshes into the second branch network, and data from intertidal mudflats into the third branch network. Each branch network performs secondary feature extraction and inversion result correction on the input data. For mangrove wetlands, the focus is on extracting spectral features of vegetation communities; for salt marshes, the focus is on extracting soil and hydrological coupling features; and for intertidal mudflats, the focus is on extracting tidal level and landform features, effectively improving the inversion accuracy of different wetland types. For wetland areas within the protected area where the sample coverage is below a preset threshold, the module completes lightweight adaptation and weight fine-tuning of the branch networks through a transfer learning adaptation interface.

[0057] In the specific implementation of this embodiment, the branch network weight fine-tuning adopts the formula Calculation. In the formula. The weights of the fine-tuned branch network are the final network weights after adapting to the target region. The pre-trained weights for the source wetland type are mature weight parameters obtained by training with enriched samples of the same type of wetland. The initial random weights for the target wetland area are initialized using a Gaussian distribution. β represents the weight fusion ratio, which is determined based on the sample coverage and ecological feature similarity of the target area. Specifically, β is set to 0.8 when the sample coverage is below 30%, 0.6 when the sample coverage is between 30% and 60%, and 0.4 when the sample coverage is above 60%. Simultaneously, the ecological feature similarity between the target area and the source area is calculated using cosine similarity. β increases by 0.1 when the similarity is above 0.8 and decreases by 0.1 when the similarity is below 0.6. In this embodiment, the final value of β for the intertidal mudflat area with scarce samples is 0.8.

[0058] After completing the inversion optimization of each branch network, the module stitches and merges the output results of the three types of branch networks according to the spatial boundaries of wetland types, eliminates regional stitching gaps, ensures the spatial continuity and integrity of the inversion results, and finally generates high-precision ecological parameter inversion results for the entire protected area. These results are then transmitted to the ecological parameter visualization and output module.

[0059] Finally, perform ecological parameter visualization accuracy assessment and data export operations, such as... Figure 1 As shown. This step is completed by the ecological parameter visualization and output module. The module has built-in thematic map generation unit, time series analysis unit, accuracy assessment unit, and data export unit, realizing comprehensive output and application of the inversion results. The thematic map generation unit receives high-precision ecological parameter inversion results and generates raster maps of spatial distribution of vegetation cover, soil salinity, and hydrological inundation depth based on the ArcGIS engine. At the same time, it generates vector thematic maps of three types of parameters. The thematic maps include complete elements such as spatial coordinates, legend, and scale, and can be directly used for wetland ecological monitoring report preparation.

[0060] The time-series analysis unit retrieves inversion results from the past three years, performs time-series variation statistics and trend fitting, analyzes the seasonal and interannual variations of ecological parameters in the protected area, and generates time-series variation curves and trend prediction reports, providing data support for ecological evolution analysis. The accuracy assessment unit verifies the inversion results based on in-situ measured data from the protected area, calculating three accuracy indicators: coefficient of determination, root mean square error, and mean absolute error. In this embodiment, the coefficients of determination for all three types of ecological parameters are higher than 0.9, and the root mean square error is lower than 5%, meeting the requirements for high-precision monitoring. The data export unit supports exporting data in various standardized formats such as TIFF, CSV, and HP, and provides an API interface for the wetland management system to call, enabling the sharing and application of monitoring data.

[0061] In summary, this embodiment achieves a complete workflow for multi-source data preprocessing, ecological mechanism knowledge base construction, knowledge-data dual-driven inversion, adaptive adaptation, and visualization output of multi-type wetlands. It transforms the coupled ecological mechanisms of coastal wetland vegetation, soil, and hydrology into differentiable constraint terms embedded in a deep learning model. By constraining the model's learning direction from both feature extraction and loss calculation dimensions, it completely avoids anomalous results from purely data-driven models that do not conform to ecological laws, significantly improving the model's interpretability and the physical rationality of the inversion results. Simultaneously, the multi-branch adaptive feature extraction network achieves differentiated adaptation for the ecological characteristics of mangrove salt marshes and intertidal mudflats. Combined with transfer learning strategies, it effectively solves the problem of insufficient inversion accuracy in areas with scarce samples, significantly improving the model's generalization ability and transfer adaptability. In actual coastal wetland monitoring scenarios, the technical solution adopted in this embodiment achieves high-precision, automated, and routine monitoring of ecological parameters of large-scale, multi-type coastal wetlands. The inversion accuracy and monitoring efficiency are improved by more than 60% compared to traditional methods, providing accurate and reliable technical support and data assurance for evaluating the effectiveness of coastal wetland protection and restoration projects and formulating ecological risk early warning management policies.

[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A deep learning-based intelligent monitoring system for coastal wetland ecological parameters, characterized in that, The system includes: a multi-source monitoring data access and preprocessing module, a coastal wetland ecological mechanism knowledge base construction module, a knowledge-data dual-driven deep learning inversion module, a multi-type wetland adaptive adaptation module, and an ecological parameter visualization and output module. The modules interact with each other through standardized data interfaces. The multi-source monitoring data access and preprocessing module accesses multi-source heterogeneous monitoring data of coastal wetlands, and after completing standardized preprocessing, generates a standardized monitoring dataset, which is then output to the coastal wetland ecological mechanism knowledge base construction module and the knowledge-data dual-driven deep learning inversion module, respectively. The coastal wetland ecological mechanism knowledge base construction module is based on the coastal wetland vegetation-soil-hydrology coupled ecological process mechanism to construct a prior constraint knowledge base, transforming the ecological mechanism rules into differentiable mechanism constraint terms, and outputting them to the knowledge-data dual-driven deep learning inversion module and the multi-type wetland adaptive adaptation module respectively. The knowledge-data dual-driven deep learning inversion module receives a standardized monitoring dataset and mechanism constraint terms, completes feature extraction, mechanism constraint embedding and end-to-end inversion of ecological parameters, and outputs the initial inversion results to the multi-type wetland adaptive adaptation module. The multi-type wetland adaptive adaptation module has a built-in multi-branch adaptive feature extraction network to complete the automatic classification of wetland types and the optimization of differential parameters, and outputs high-precision ecological parameter inversion results to the ecological parameter visualization and output module. The ecological parameter visualization and output module completes the visualization display, accuracy assessment and data export of the inversion results.

2. The intelligent monitoring system for coastal wetland ecological parameters based on deep learning according to claim 1, characterized in that, The multi-source monitoring data access and preprocessing module performs the following operations during data access and preprocessing: It integrates satellite multispectral remote sensing image data, satellite hyperspectral remote sensing image data, UAV aerial survey image data, UAV LiDAR point cloud data, in-situ hydrological monitoring data, in-situ soil monitoring data, in-situ vegetation monitoring data, and tidal hydrological observation data, and completes the format unification and storage of various types of data. For optical image data, radiometric calibration, atmospheric correction, geometric fine correction, and noise removal are performed; for in-situ site data, outlier removal, missing value interpolation, and time-series alignment are performed. Spatial registration and data normalization are performed on all preprocessed data to generate a standardized monitoring dataset with matching spatiotemporal dimensions.

3. The intelligent monitoring system for coastal wetland ecological parameters based on deep learning according to claim 1, characterized in that, The prior constraint knowledge base constructed by the coastal wetland ecological mechanism knowledge base construction module includes the mapping relationship rules between vegetation spectral response characteristics and physiological and ecological parameters corresponding to different coastal wetland types, the mapping relationship rules between vegetation spectral response characteristics and environmental driving factors, the ecological constraint thresholds of core ecological parameters under different tidal conditions, and the spatiotemporal variation patterns of core ecological parameters under different seasonal cycles. The module transforms all ecological mechanism rules in the prior constraint knowledge base into differentiable mechanism constraint terms that can be embedded in deep learning networks. The differentiable mechanism constraint terms are divided into two categories: feature layer constraint terms and loss layer constraint terms.

4. The intelligent monitoring system for coastal wetland ecological parameters based on deep learning according to claim 3, characterized in that, The coastal wetland ecological mechanism knowledge base construction module performs the following operations when constructing the prior constraint knowledge base and mechanism constraint items: By integrating the mechanistic research results of the coupled ecological processes of vegetation-soil-hydrology in coastal wetlands with long-term in-situ observation data, the mapping relationship rules between vegetation spectral response characteristics and physiological and ecological parameters and environmental driving factors are extracted. Based on the differences in ecological characteristics of different types of coastal wetlands, corresponding ecological units are divided, the ecological constraint thresholds and spatiotemporal variation patterns of core ecological parameters within each ecological unit are determined, and a priori constraint knowledge base is constructed. The mapping rules and constraint thresholds in the prior constraint knowledge base are transformed into differentiable mathematical expressions to generate feature layer constraint terms and loss layer constraint terms. The loss layer constraint terms are calculated using the following formula: ,in: The ecological mechanism constraint loss value is N, where N is the total number of ecological parameter types. The mechanistic constraint weights for the i-th type of ecological parameter are determined based on the statistical characteristics of long-term ecological observation data of coastal wetlands. The values ​​of the i-th type of ecological parameters retrieved from the model are as follows: The ecological mechanism threshold of the i-th type of ecological parameter in the prior constraint knowledge base; The generated two types of mechanism constraint terms are format-standardized and output to the knowledge-data dual-driven deep learning inversion module.

5. The intelligent monitoring system for coastal wetland ecological parameters based on deep learning according to claim 1, characterized in that, The knowledge-data dual-driven deep learning inversion module includes a feature extraction submodule, a mechanism constraint embedding submodule, and an end-to-end parameter inversion submodule. The feature extraction submodule adopts a multi-level convolutional neural network architecture. The mechanism constraint embedding submodule is connected to the coastal wetland ecological mechanism knowledge base construction module through a standardized data interface. The end-to-end parameter inversion submodule adopts a multi-branch decoding structure, and the number of branches in the multi-branch decoding structure is consistent with the number of ecological parameter types to be inverted.

6. The intelligent monitoring system for coastal wetland ecological parameters based on deep learning according to claim 5, characterized in that, The knowledge-data dual-driven deep learning inversion module performs the following operations during feature extraction and parameter inversion: The feature extraction submodule receives a standardized monitoring dataset and extracts spectral features, texture features, and spatiotemporal correlation features of the input data through a multi-level convolutional neural network to generate a multi-level initial feature map. The mechanism constraint embedding submodule receives differentiable mechanism constraint terms, embeds the feature layer constraint terms into the multi-level feature extraction stage of the feature extraction submodule, performs constraint optimization on the multi-level initial feature map, and generates the optimized feature map; The mechanism constraint embedding submodule weights and fuses the loss layer constraint terms with the data fitting loss terms to construct the joint loss function of the model. The joint loss function is calculated using the following formula: ,in, This represents the total loss value of the model. Fit loss values ​​to the model data. The ecological mechanism constraint loss value is α, which is the loss fusion coefficient. This coefficient is determined through adaptive iterative optimization of the ecological parameter inversion accuracy during model training. The end-to-end parameter inversion submodule receives the optimized feature map, completes feature decoding and spatial dimension restoration through a multi-branch decoding structure, and outputs the initial inversion results corresponding to various ecological parameters.

7. The intelligent monitoring system for coastal wetland ecological parameters based on deep learning according to claim 1, characterized in that, The multi-branch adaptive feature extraction network built into the multi-type wetland adaptive adaptation module includes a first branch network corresponding to mangrove wetlands, a second branch network corresponding to salt marsh wetlands, and a third branch network corresponding to intertidal mudflats. Each branch network adopts independent convolutional kernel size, network layer number, and feature weight parameters. The structural parameters of each branch network are matched with the ecological features of the corresponding wetland type, and each branch network is set with an independent transfer learning adaptation interface.

8. The intelligent monitoring system for coastal wetland ecological parameters based on deep learning according to claim 7, characterized in that, The multi-type wetland adaptive adaptation module performs the following operations when classifying wetland types and optimizing parameters inversion: Receive standardized monitoring datasets and initial inversion results, and based on the wetland type classification rules in the prior constraint knowledge base, complete the automatic identification and spatial boundary delineation of mangrove wetlands, salt marsh wetlands, and intertidal mudflats within the monitoring area, and generate regional masks corresponding to each wetland type; Based on the regional mask of each wetland type, the optimized feature map of the corresponding region and the initial inversion result are input into the matching branch network to complete the secondary extraction of the corresponding region features and the correction of the inversion result; For wetland areas where sample coverage is below a preset threshold, lightweight adaptation and weight fine-tuning of the corresponding branch network are completed through the transfer learning adaptation interface. The weight fine-tuning is calculated using the following formula: ,in, To fine-tune the branch network weights, For the pre-trained weights of the source wetland type, β represents the initial random weights for the target wetland area, and β is the weight fusion ratio, which is calculated and determined based on the sample coverage and ecological feature similarity of the target area. The inversion results output from each branch network are spliced ​​and merged according to the spatial boundaries to generate high-precision ecological parameter inversion results for the entire monitoring area.

9. The intelligent monitoring system for coastal wetland ecological parameters based on deep learning according to claim 1, characterized in that, The ecological parameter visualization and output module includes a thematic map generation unit, a time series analysis unit, an accuracy evaluation unit, and a data export unit. The thematic map generation unit receives high-precision ecological parameter inversion results and generates spatial distribution raster maps and vector thematic maps corresponding to various ecological parameters. The time series analysis unit performs time series variation statistics and trend fitting of multiple inversion results. The accuracy evaluation unit performs error statistics and accuracy verification of the inversion results based on in-situ measured data. The data export unit supports exporting inversion data in various standardized formats and calling external interfaces.

10. The intelligent monitoring system for coastal wetland ecological parameters based on deep learning according to claim 1, characterized in that, The output of the multi-source monitoring data access and preprocessing module is connected to the input of the coastal wetland ecological mechanism knowledge base construction module and the input of the knowledge-data dual-driven deep learning inversion module. The output of the coastal wetland ecological mechanism knowledge base construction module is connected to the input of the knowledge-data dual-driven deep learning inversion module and the input of the multi-type wetland adaptive adaptation module. The output of the knowledge-data dual-driven deep learning inversion module is connected to the input of the multi-type wetland adaptive adaptation module. The output of the multi-type wetland adaptive adaptation module is connected to the input of the ecological parameter visualization and output module. The standardized data interfaces between the modules adopt a unified communication protocol and data format specification.

Images