Wetland plant diversity monitoring device and monitoring system

Abstract

The invention relates to the technical field of wetland plant diversity monitoring devices, and particularly discloses a wetland plant diversity monitoring device and monitoring system, which comprises a multi-source data acquisition unit, a data transmission unit and a data encryption unit, the multi-source data acquisition unit comprises an air acquisition unit, a ground acquisition unit and an underwater acquisition unit; the air acquisition unit comprises a multi-rotor unmanned aerial vehicle, and a hyperspectral imager and an RGB (Red, Green, Blue) high-definition camera which are carried on the multi-rotor unmanned aerial vehicle; according to the invention, on one hand, through the multi-source acquisition unit, the limitation that areas such as deep water areas and marshes are difficult to cover by manual monitoring is broken through, and the whole habitat coverage of emergent water, floating water and submerged plants is realized; and on the other hand, the defect of single dimension of existing automatic monitoring data is made up by collaborative acquisition of multi-source data.

Description

Technical Field

[0001] This invention relates to the field of wetland plant diversity monitoring devices, and in particular to a wetland plant diversity monitoring device and monitoring system. Background Technology

[0002] Wetlands, as unique ecosystems where water and land intersect, are one of the core carriers of global biodiversity. Changes in the structure and diversity of their plant communities directly reflect the health of the ecosystem and play an irreplaceable role in conserving water resources, purifying water quality, and maintaining regional ecological balance. Currently, wetland plant diversity monitoring still relies on traditional manual survey methods, such as setting up quadrats, collecting specimens, visually identifying species, and recording data. This approach not only has limitations in coverage—deep water areas and marshes, which are difficult for personnel to access, cannot be monitored—but also suffers from low efficiency, human intervention disturbing the wetland ecosystem, and reliance on experience for species identification leading to significant subjective errors. Therefore, it is difficult to meet the needs of dynamic wetland ecological monitoring.

[0003] While existing automated monitoring technologies have partially replaced manual operations, they are mostly based on single sensors for local monitoring, resulting in limited data dimensions and a lack of effective fusion and dynamic calibration mechanisms for multi-source data, leading to significant deviations in monitoring data. Summary of the Invention

[0004] In response to the technical problems mentioned in the background art, the present invention provides a wetland plant diversity monitoring device and monitoring system.

[0005] The technical solution adopted in this invention is: a wetland plant diversity monitoring device, comprising a multi-source data acquisition unit, a data transmission unit, and a data encryption unit; the multi-source data acquisition unit includes an aerial acquisition unit, a ground acquisition unit, and an underwater acquisition unit; the aerial acquisition unit includes a multi-rotor drone, and a hyperspectral imager and an RGB high-definition camera mounted on the multi-rotor drone; the ground acquisition unit includes a sensor array and a portable hyperspectral imager deployed in typical wetland sample plots, the sensor array including a soil temperature and humidity sensor, a light sensor, and a water level sensor; the underwater acquisition unit includes an underwater high-definition camera and an underwater spectral sensor deployed at different water depth gradients in the wetland; the data transmission unit adopts a 5G and LoRa dual-mode communication architecture for transmitting the data acquired by the multi-source data acquisition unit to a backend server;

[0006] The data encryption unit uses the AES-128 encryption algorithm to encrypt data before transmission.

[0007] The invention is further configured to include a wetland plant diversity monitoring device, a data preprocessing and multi-source data fusion module, a wetland plant species intelligent identification module, a plant diversity core indicator calculation module, a wetland plant diversity dynamic analysis module, a diversity anomaly early warning and protection suggestion generation module, a monitoring data visualization and interaction module, and a system adaptive calibration and maintenance module. The data preprocessing and multi-source data fusion module performs noise reduction, correction, and registration preprocessing on the raw data, calculates the weights of each data source using the entropy weight method, and then completes multi-source data fusion based on a weighted fusion formula to generate a standardized dataset. The wetland plant species intelligent identification module uses an improved ResNet-50 model to identify emergent, floating, and submerged plants in the wetland. The plant diversity core indicator calculation module calculates the Shannon-Wiener index and Jaccard similarity index with introduced species dominance weights. The wetland plant diversity dynamic analysis module analyzes the temporal trend and spatial distribution characteristics of diversity indicators. The diversity anomaly early warning and protection suggestion generation module generates tiered early warnings based on trend slope calculation results and dynamic thresholds, and automatically generates protection suggestions.

[0008] The monitoring data visualization and interaction module is used to realize spatial visualization, temporal visualization and early warning visualization of monitoring data; the system adaptive calibration and maintenance module is used to complete sensor error correction and recognition model accuracy calibration.

[0009] The present invention is further configured such that the data preprocessing and multi-source data fusion module performs the following specific processing steps:

[0010] Data preprocessing: Radiometric, atmospheric, and geometric corrections are performed on hyperspectral images from multi-rotor UAVs; noise reduction and brightness equalization are performed on RGB and underwater images;

[0011] Sensor data preprocessing: Outlier removal and missing value completion are performed on time-series data such as soil temperature and humidity, light intensity, and water level, and the data is then standardized.

[0012] After multi-source data fusion preprocessing, hyperspectral data, RGB morphological data, and environmental sensor data need to be weighted and fused. The entropy weight method is used to calculate the weight of each data source. The weight calculation formula is as follows:

[0013] ;

[0014] in, For the first The weight coefficients of each data source must satisfy... For the first The information entropy of each data source is calculated as follows: The number of data sources to be merged; For the first The number of samples from each data source; For the first The first data source Normalized frequencies of each sample;

[0015] Based on the theory of information entropy, the lower the information entropy of a data source, the lower its uncertainty and the higher its effective information content, and the larger its corresponding weight coefficient. After determining the weights, the following formula is used to complete the fusion of multi-source data and generate a standardized dataset containing three-dimensional information:

[0016] ;

[0017] in, For the merged standardized data, For the first Standardized data from a data source after preprocessing.

[0018] A further provision of the present invention is that the specific process of the wetland plant species intelligent identification module is as follows:

[0019] Establish a wetland plant sample bank: Collect hyperspectral data, morphological feature data, and growth environment data of common plant species in the target wetland to construct an annotated sample bank covering different growth stages and environmental conditions; Improved CNN recognition model: Use an improved ResNet-50 model as the base network, add a multi-scale feature fusion layer at the model input to fuse the spectral features of hyperspectral data and the morphological features of RGB data; Introduce an attention mechanism in the intermediate layers of the model to enhance the model's ability to extract key plant features; Use the Softmax function in the output layer to achieve multi-classification, outputting species recognition results and confidence scores; Model optimization: To address the problem of low accuracy in identifying submerged plants due to dim underwater lighting in wetlands, add underwater environment-enhanced samples to the sample bank and incorporate an underwater environment adaptive loss function during model training.

[0020] A further provision of the present invention specifies the following process for the calculation module of the core indicators of plant diversity:

[0021] Basic data statistics: Based on the species identification results, the number of species and the number of individuals of each species in each monitoring plot are counted;

[0022] Diversity index calculation: To address the shortcomings of the traditional Shannon-Wiener index, which only considers relative species abundance and ignores the impact of dominant species on the community, a species dominance weight is introduced. The improved Shannon-Wiener index calculation formula is as follows:

[0023] ;

[0024] in, The Shannon-Wiener index is an improved version; a higher value indicates greater species diversity. This represents the total number of species within the sample plot; For the first The relative abundance of each species; For the first The method for calculating the dominance weight of each species is as follows: For the first Average plant height of each species For the first Coverage of individual species;

[0025] To analyze species similarity across different habitat regions, the Jaccard similarity index in β-diversity is calculated using the following formula: in, The Jaccard similarity index has a value range of [0,1]. The larger the value, the higher the species similarity between the two plots. The number of species shared by the two plots; The number of species that appear only in the first plot; This represents the number of species that appeared only in the second plot;

[0026] By combining information such as plot area and species distribution, we can calculate species richness, Simpson index and other diversity indicators. For the habitat differences in different wetland areas, we can calculate the diversity indicators in each area to achieve zoning analysis.

[0027] A further provision of the present invention specifies the following process for the wetland plant diversity dynamic analysis module:

[0028] Time series analysis: Trend fitting and abrupt change test methods were used to analyze the temporal trends of diversity indicators. Monthly, quarterly, and annual diversity indicator change curves were fitted using linear regression, and the Mann-Kendall test was used to detect abrupt changes in diversity and determine the significance of the trends. Spatial distribution analysis: Kriging interpolation was used to interpolate the diversity indicators of each monitoring plot into a continuous spatial distribution layer, visually displaying the diversity levels of different wetland areas. Combined with GIS technology, wetland habitat type maps were overlaid to analyze diversity differences under different habitat types. Correlation analysis: Redundancy analysis was used to analyze the correlation between diversity indicators and environmental factors, identifying key environmental factors affecting wetland plant diversity.

[0029] A further provision of the present invention specifies the following process for the diversity anomaly early warning and protection suggestion generation module:

[0030] Dynamic trend assessment: Based on the time series analysis results, the trend of diversity changes is determined using the linear regression slope formula, which is as follows:

[0031] ;

[0032] in, The trend slope This indicates an upward trend in diversity. This indicates a downward trend. The larger the value, the more pronounced the trend. To monitor the length of the time series; It is a time variable; For the first Shannon-Wiener index at each time point;

[0033] Anomaly threshold setting and early warning determination: The mean ± 2 standard deviation of 3 years of historical monitoring data is used as the basic threshold. Combined with wetland ecological protection goals, and revised through expert review, dynamic anomaly thresholds for different regions and seasons are determined. Based on the dynamic thresholds, the following formula is used to determine the early warning level:

[0034] ;

[0035] in, The Shannon-Wiener index is currently being monitored. The abnormal threshold is dynamically set; if the indicator exceeds the threshold range, the system will automatically trigger an alert, and the alert information will be pushed to the management personnel via SMS and APP.

[0036] Protection Recommendation Generation: Based on the results of correlation analysis, key environmental factors leading to abnormal diversity are identified. Combined with the ecological habits of wetland plants, a pre-set knowledge base of protection measures is invoked to automatically generate targeted protection recommendations.

[0037] The present invention is further configured such that the monitoring data visualization and interaction module operates as follows: Visualization display: A web-based visualization platform is used, and the displayed content includes: Spatial visualization: a species distribution map of the entire wetland area and a spatial distribution layer of diversity indicators, supporting zooming, panning, and click-to-query; Temporal visualization: monthly, quarterly, and annual trend charts of diversity indicators, supporting comparative display of different indicators and different regions; Early warning visualization: highlighted markings of abnormal areas and early warning level icons, with clicks on early warning areas allowing viewing of early warning details and protection recommendations;

[0038] Statistical Reports: Automatically generates monthly, quarterly, and annual monitoring reports, including data statistics, trend analysis, early warning information, and protection recommendations.

[0039] Interactive functions: Provides data query, data export, spatial positioning, and parameter setting functions.

[0040] A further provision of the present invention specifies the following process for the system adaptive calibration and maintenance module:

[0041] Sensor calibration: On-site calibration of ground and underwater sensors is performed quarterly. Sensor measurements are compared using standard equipment, and systematic errors are eliminated using an error correction formula:

[0042] ;

[0043] in, The corrected sensor measurement value. These are the original measurements from the sensor; The sensor error is calculated as follows: ;

[0044] Model calibration: New plant samples are collected every six months to supplement the sample bank. The improved CNN model is then retrained and fine-tuned to enhance its ability to identify new species or new growth stages of species. Combining the original model validation results with the results of manual field verification, a weighted formula is used to calibrate the model's accuracy. The formula is as follows:

[0045] ;

[0046] in, The accuracy of the calibrated model is the recognition rate. This represents the model's original recognition accuracy. To improve the accuracy of manual on-site verification, The weighting coefficients are used to calculate the model's recognition accuracy in real time. If the accuracy falls below 90%, an emergency calibration process is automatically triggered.

[0047] Algorithm parameter calibration: Based on long-term monitoring data, the weight coefficients of the data fusion algorithm and the parameters of the diversity analysis algorithm are regularly optimized to make the algorithm more adaptable to changes in the wetland environment; an algorithm error assessment mechanism is established, through manual on-site verification of data and comparison with system monitoring data, errors are calculated and parameters are adjusted;

[0048] System maintenance: Establish an equipment status monitoring mechanism to monitor the operating status of sensors, multi-rotor drones, and servers in real time. If equipment failure occurs, an automatic maintenance reminder will be issued. The database will be backed up and optimized regularly.

[0049] The beneficial effects of this invention are:

[0050] Firstly, this invention overcomes the limitations of manual monitoring in covering deep water areas and marshes through multi-source acquisition units, achieving full habitat coverage of emergent, floating, and submerged plants. It also avoids the trampling disturbance to wetland ecosystems caused by manual on-site sampling, replacing experience-based manual species identification and significantly reducing subjective errors. Secondly, the collaborative acquisition of multi-source data compensates for the single-dimensionality of existing automated monitoring data, making subsequent species identification more adaptable to the complex wetland environment. Combined with encrypted dual-mode transmission, it ensures stable and secure data transmission, solving the problems of delayed and fragmented data in traditional monitoring.

[0051] Second, this invention fills the gap in submerged plant monitoring through a multi-source data collaborative acquisition module, achieving comprehensive, blind-spot-free monitoring of emergent, floating, and submerged plants. Dual-mode communication balances efficient transmission with low power consumption, reducing human intervention and interference with the wetland ecosystem. The collected multi-dimensional data lays a solid foundation for subsequent analysis. The data preprocessing and multi-source data fusion module specifically eliminates environmental interference such as atmospheric scattering and water refraction, as well as equipment errors. It enhances the distinction between plants and the background through differentiated optimization strategies and uses the objective entropy weight method to organically integrate multi-source data, generating a standardized fusion dataset. This overcomes the limitations of single data sources, achieves data value-added, and provides high-quality input for subsequent analysis. The intelligent wetland plant species identification module, with its dedicated wetland plant sample library, adapts to complex monitoring scenarios. It enhances key feature extraction through model structure improvements and addresses the pain points of submerged plant identification through sample augmentation technology. The confidence judgment mechanism balances efficiency and accuracy. The core plant diversity indicator calculation module improves upon traditional indices by introducing dominance weights to align with the realities of wetland communities. It employs refined zoning calculations to reveal differences in diversity across different habitats, achieving a scientific evaluation from qualitative to quantitative perspectives and providing a core basis for dynamic analysis. The wetland plant diversity dynamic analysis module conducts temporal and spatial analysis, accurately revealing trends, identifying key nodes and regional differences, and deeply exploring the intrinsic relationship between diversity and environmental factors. This overcomes the limitations of static analysis and provides a scientific basis for early warning and decision-making. The diversity anomaly early warning and protection suggestion generation module uses dynamic thresholds to adapt to environmental changes, a tiered early warning mechanism to accurately convey the degree of anomaly, and multi-channel push notifications to ensure timely response. The monitoring data visualization and interaction module transforms complex data into intuitive visualizations of spatial distribution and temporal trends. It supports multi-condition queries, data export, and parameter settings, and combines RBAC permission management to ensure data security, significantly improving the system's practicality and ease of use, allowing professional data to efficiently serve management decisions. The system adaptive calibration and maintenance module establishes a routine calibration and maintenance mechanism, regularly correcting sensor errors, optimizing recognition models and algorithm parameters, monitoring equipment status in real time, promptly troubleshooting, ensuring long-term stable system operation, and guaranteeing continuously accurate and reliable monitoring data. Attached Figure Description

[0052] Figure 1This is a schematic diagram of the structure of the multi-rotor UAV of the present invention;

[0053] Figure 2 This is an example diagram of the sensor array deployed using a grid method in the ground acquisition unit of this invention.

[0054] The diagram is marked as follows:

[0055] 1. Multi-rotor drone; 2. RGB high-definition camera; 3. Hyperspectral imager. Detailed Implementation

[0056] In the description of this invention, it should be noted that the terms "front", "up", "down", "left", "right", "vertical", "horizontal", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0057] The following is in conjunction with the appendix Figure 1-2 The present invention will be further described below.

[0058] To address the problems existing in the background technology, this application proposes the following technical solution: a wetland plant diversity monitoring device, comprising a multi-source data acquisition unit, a data transmission unit, and a data encryption unit;

[0059] The multi-source data acquisition unit includes an aerial acquisition unit, a ground acquisition unit, and an underwater acquisition unit;

[0060] The aerial acquisition unit includes a multi-rotor drone 1, and a hyperspectral imager 3 and an RGB high-definition camera 2 mounted on the multi-rotor drone 1.

[0061] The ground acquisition unit includes a sensor array and a portable hyperspectral analyzer deployed in typical wetland sample plots. The sensor array includes a soil temperature and humidity sensor, a light sensor, and a water level sensor.

[0062] The underwater acquisition unit includes underwater high-definition cameras and underwater spectral sensors deployed at different water depth gradients in the wetland.

[0063] The data transmission unit adopts a 5G and LoRa dual-mode communication architecture to transmit the data collected by the multi-source data acquisition unit to the backend server.

[0064] The data encryption unit uses the AES-128 encryption algorithm to encrypt data before transmission.

[0065] Technical solution:

[0066] Aerial data acquisition unit: A multi-rotor UAV 1 equipped with a hyperspectral imager 3 (spectral range 400-1000nm, spectral resolution 5nm) and an RGB high-definition camera 2 is used to fly along a preset route (along wetland contour lines, flight altitude 50m, overlap rate 80%) to acquire hyperspectral and RGB images of the entire wetland area and capture the macroscopic distribution and spectral characteristics of emergent and floating plants.

[0067] Equipment selection: Multi-rotor drone 1 (model: DJI Matrice 350RTK, maximum takeoff weight 9.6kg, flight time 38 minutes), equipped with hyperspectral imager 3 (model: Headwall Nano-Hyperspec, spectral range 400-1000nm, resolution 5nm, pixel 320×256) and RGB high-definition camera 2 (model: Zenmuse P1, full-frame 24-megapixel, shutter speed 1 / 100-1 / 1000s);

[0068] Flight route planning: Based on a 1:1000 precision wetland GIS vector map, the "parallel flight of contour lines" mode is adopted, with a flight altitude of 50m (ground resolution 0.1m / pixel), a speed of 6m / s, a forward overlap rate of 80%, and a lateral overlap rate of 70% (to avoid image stitching defects).

[0069] Collection schedule: During the non-flood season (April-May, September-October), twice a month (15 days apart); during the flood season (June-August), once a month (avoiding the 24 hours after a rainstorm); and during the winter season (November-March), once every two months. Collection time is 10:00-14:00 (when the light is stable and shadows are minimized).

[0070] Ground-based data collection unit: Sensor arrays are deployed in typical wetland plots (laid out using a grid method, with a grid size of 10m×10m, and densified to 5m×5m in the core area). These arrays include soil temperature and humidity sensors (measurement range 0-100%RH, accuracy ±2%RH), light sensors (measurement range 0-200000lux, accuracy ±5%), and water level sensors (measurement range 0-3m, accuracy ±1cm) to collect environmental parameters for plant growth in real time. Simultaneously, fixed monitoring points are set up in the plots, equipped with portable hyperspectral analyzers, to collect leaf spectral samples of dominant plants periodically (once a month) for subsequent model calibration.

[0071] Sample plot layout: The plots are laid out according to the "grid method and habitat stratification". For wetlands ≤100hm², the grid is 10m×10m. For wetlands >100hm², the non-core area is 10m×10m and the core area (around bird habitats) is densified to 5m×5m. The total number of sample plots is ≥30, and the edge sample plots are ≥50m away from the wetland boundary.

[0072] Sensor selection: Soil temperature and humidity sensor (SensirionSHT30, measurement range 0-100%RH, -40-125℃, accuracy ±2%RH, ±0.3℃), light sensor (TSL2591, 0-200000lux, accuracy ±5%), water level sensor (DYP-A01F, 0-3m, accuracy ±1cm).

[0073] Deployment and data collection: The sensors were fixed with stainless steel brackets. The soil temperature and humidity sensor was inserted into the soil 10cm (dense root layer), and the water level sensor was fixed at the lowest point of the plot. The data collection frequency was 1 hour / time. The portable hyperspectral analyzer (ASDFieldSpec4, 350-2500nm) collected the leaf spectra of the dominant plants on the 15th of each month (3 samples per species, 3 measurements per sample and average).

[0074] Underwater acquisition unit: Equipped with an underwater high-definition camera and an underwater spectral sensor in the wetland water area (deployed according to the water depth gradient, one unit each at water depths of 0.5m, 1m, and 2m), it captures morphological images of submerged plants in real time and collects underwater illumination and water spectral data, making up for the shortcomings of traditional monitoring that ignore submerged plants.

[0075] Deployment principle: Deployed according to water depth gradient (0.5m, 1m, 2m), with 3 monitoring points for each gradient (distributed in an equilateral triangle with a side length of 5m), covering deep water area, shallow water area, and emergent-submerged transition zone;

[0076] Equipment selection: underwater HD camera (Hikvision DS-2CD3146WD-I, IP68 waterproof, 1080P resolution, minimum illumination 0.01 lux), underwater spectral sensor (JAZ-COM-2, 200-850nm, resolution 0.35nm).

[0077] Fixation and data collection: The anchor bracket is fixed to the bottom of the water, the camera lens faces upward (to capture the canopy of submerged plants), and the sensor is 30cm away from the bottom of the water (to avoid interference from silt); data is collected every 2 hours, and video is collected every 10 minutes (to reduce storage pressure).

[0078] Data transmission: Adopting dual-mode communication of "5G and LoRa", the data of the multi-rotor UAV1 is transmitted back in real time via 5G, and the ground and underwater sensor data is transmitted to the edge gateway via LoRa low power, and then aggregated to the back-end server to ensure the stability and real-time performance of data transmission.

[0079] Implementation details:

[0080] Preparations before deployment: First, delineate wetland boundaries and habitat types (deep water area / shallow water area / tidal flat area / buffer zone) using 0.5m resolution satellite remote sensing images, and then conduct on-site surveys of sample plot feasibility (avoiding soft mud areas and wildlife nests);

[0081] Equipment debugging: Before the first flight of the multi-rotor UAV1, the magnetic compass and IMU were calibrated, and the image resolution was verified by a 1m×1m black and white target; after the sensor was installed, data was collected continuously for 24 hours, and outliers were removed using the 3σ criterion to ensure data stability.

[0082] The aforementioned technical solution, by constructing a three-in-one data acquisition architecture encompassing aerial, ground, and underwater sensors, completely solves the core problem of limited coverage in traditional monitoring. It achieves comprehensive, blind-spot-free monitoring of wetland plants, particularly filling the gap in submerged plant monitoring, allowing for the complete capture of the growth status of emergent, floating, and submerged plants. The aerial acquisition unit can quickly obtain macroscopic vegetation distribution information of the wetland, the ground acquisition unit accurately collects environmental parameters for plant growth, and the underwater acquisition unit overcomes the technical barriers of water area monitoring. The dual-mode communication method balances the high efficiency and low power consumption requirements of data transmission, ensuring stable data transmission. The entire acquisition process requires minimal on-site human intervention, reducing human interference with the wetland ecosystem. Simultaneously, the collected multi-dimensional data provides comprehensive and solid foundational support for subsequent species identification and diversity analysis, improving the scientific rigor and reliability of monitoring work from the source.

[0083] One of them is a wetland plant diversity monitoring system, which includes a wetland plant diversity monitoring device, a data preprocessing and multi-source data fusion module, a wetland plant species intelligent identification module, a plant diversity core indicator calculation module, a wetland plant diversity dynamic analysis module, a diversity anomaly early warning and protection suggestion generation module, a monitoring data visualization and interaction module, and a system adaptive calibration and maintenance module.

[0084] Data preprocessing and multi-source data fusion module: Preprocesses the collected raw data by noise reduction, correction, registration, etc., to eliminate environmental interference and equipment errors. Then, it integrates data from different dimensions through multi-source data fusion technology to generate a standardized monitoring dataset, laying the foundation for subsequent species identification and diversity analysis.

[0085] Technical solution:

[0086] Data preprocessing:

[0087] Radiometric correction (eliminating sensor response differences), atmospheric correction (using the FLAASH model to eliminate the influence of atmospheric scattering), and geometric correction (using GPS control points as a reference, with a correction error ≤ 1 pixel) were performed on the hyperspectral images of the multi-rotor UAV 1; noise reduction (using an adaptive median filtering algorithm) and brightness equalization were performed on the RGB images and underwater images to enhance the contrast between plants and the background.

[0088] The details of data preprocessing are as follows:

[0089] Hyperspectral image preprocessing:

[0090] Radiation correction: Dark current correction file using sensors, formula. ( Corrected radiance Original value, Dark current value);

[0091] Atmospheric correction: FLAASH model, parameters set to "mid-latitude seasonal matching" and "rural aerosols", water vapor content retrieved through the 940mm absorption band, correction error ≤5%;

[0092] Geometric correction: Using GPS control points with an accuracy of ±2cm as a reference, quadratic polynomial interpolation method is used to correct the error. oRGB / Underwater Image Preprocessing:

[0093] Noise reduction: Adaptive median filtering, window size 3×3-7×7 (7×7 for cloudy / rainy days, 3×3 for sunny days);

[0094] Brightness equalization: Histogram equalization stretches grayscale values ​​to 0-255; underwater images receive additional blue-green channel enhancement, formula... ( Enhanced green channel value, Original green channel value, Original blue channel value;

[0095] Sensor data preprocessing: outlier removal (using the 3σ criterion) and missing value completion (using linear interpolation) are performed on time-series data such as soil temperature and humidity, light intensity, and water level, and the data is standardized (mapped to the [0,1] interval).

[0096] The details of sensor time-series data preprocessing are as follows:

[0097] Outlier Removal: 3σ Criterion (Removing outliers) (Check sensor malfunctions when the percentage of outliers in the data is >5%).

[0098] Missing value completion: Linear interpolation method, formula ( Missing values, (Missing moment);

[0099] Standardization: min-max standardization maps to [0,1], formula

[0100] After multi-source data fusion preprocessing, hyperspectral data, RGB morphological data, and environmental sensor data need to be weighted and fused. To ensure the objectivity of the fusion weights, the entropy weight method is used to calculate the weights of each data source. The weight calculation formula is as follows:

[0101] ;

[0102] in, For the first The weight coefficients of each data source must satisfy... For the first The information entropy of each data source is calculated as follows: The number of data sources to be merged (in this system) ); For the first The number of samples from each data source; For the first The first data source Normalized frequencies of each sample.

[0103] The details of multi-source data fusion are as follows:

[0104] Fusion data source: preprocessed hyperspectral data ( ), RGB morphological data ( ), environmental sensor data ( );

[0105] Weight calculation: Entropy weighting method (objective weighting), the steps are as follows: ( The standardized value of the j-th sample of the i-th data type (m samples); calculate the information entropy. ( The smaller the value, the more effective information is available; calculate the weight. (satisfy Weighted fusion: formula Fusion verification: Calculate the signal-to-noise ratio (SNR≥30dB is acceptable, otherwise reprocess).

[0106] Based on information entropy theory, data sources with lower information entropy have lower uncertainty, higher effective information content, and correspondingly larger weighting coefficients. After determining the weights, multi-source data fusion is performed using the following formula to generate a standardized dataset containing three-dimensional information of "spectral-morphological-environment":

[0107] ;

[0108] in, For the merged standardized data, For the first Standardized data from a data source after preprocessing.

[0109] Implementation details: Data preprocessing is implemented using Python open-source libraries (such as GDAL for image correction and Pandas for time series data processing); the fusion algorithm is deployed on the edge gateway to reduce data transmission pressure; the monitoring data of the previous day is automatically preprocessed and fused every morning to generate a standardized dataset and store it in the database.

[0110] Preprocessing tool details: Python 3.9, dependent on GDAL 3.6 (image correction), OpenCV 4.8 (denoising), and Pandas 1.5 (time series data processing). The algorithm is deployed on an Intel NUC12 Pro edge gateway. 2. Data storage: The fused data is stored in a PostgreSQL 14 database, stored in layers according to "date-region-data source" (e.g., 20240515-deep water area-hyperspectral), retaining 5 years of data and supporting fast querying.

[0111] In the above technical solution: This module addresses environmental interference and equipment errors in the original acquired data by performing targeted preprocessing, effectively eliminating the negative impacts of atmospheric scattering, water refraction, and sensor drift, significantly improving data purity. Differentiated optimization strategies are employed for different types of image data, enhancing the distinction between plants and background, making subsequent feature extraction more accurate. Outlier removal and missing value completion in time-series sensor data ensure the continuity and validity of environmental parameters. The multi-source data fusion stage uses an objective weight calculation method to avoid subjective weighting bias, organically integrating spectral, morphological, and environmental data to generate a standardized fusion dataset. This dataset simultaneously possesses both plant external characteristics and growth environment information, breaking the information limitations of a single data source, providing high-quality input data for subsequent species identification models, and significantly improving the accuracy and reliability of subsequent analysis.

[0112] Intelligent identification module for wetland plant species: Based on a fusion dataset, it can accurately identify emergent, floating, and submerged plants in wetlands, and output information such as species name, quantity, and distribution location, solving the problems of low efficiency and low accuracy of traditional manual identification and poor adaptability of existing automated identification models.

[0113] Technical solution:

[0114] Establish a wetland plant sample bank: Collect hyperspectral data, morphological characteristic data (leaf shape, flower color, plant height) and growth environment data of common plant species in the target wetland (such as reed, cattail, lotus, and foxtail grass), and construct a labeled sample bank containing 10,000+ samples, covering different growth stages (seedling stage, growth stage, and maturity stage) and different environmental conditions (different water levels and light).

[0115] Details of the construction of the wetland plant sample bank:

[0116] Sample scope: Covering the target wetland and more than 30 plant species within a 5km radius (15 emergent species, 5 floating species, and 10 submerged species, such as reeds, cattails, and foxtail grass).

[0117] Sample criteria: 1000+ samples per species, covering the seedling stage (200+), growth stage (500+), maturity stage (300+), and different water conditions (sunny / cloudy / rainy days); hyperspectral samples are 320×256 pixels, RGB samples are 2400×1600 pixels, and underwater samples are 1920×1080 pixels.

[0118] Labeling specifications: LabelImg tool is used for labeling, including species name, leaf / canopy outline, and collection information. The outline error is ≤2 pixels, and the species labeling accuracy is 100% (cross-reviewed by 2 botanical experts).

[0119] Sample enhancement: Underwater samples were augmented to 500+ per species by adjusting brightness (±30%), contrast (±20%), and adding 3×3-5×5 Gaussian blur.

[0120] Improved CNN recognition model: An improved ResNet-50 model is used as the base network. A multi-scale feature fusion layer is added at the model input to fuse the spectral features of hyperspectral data and the morphological features of RGB data. An attention mechanism (CBAM module) is introduced in the middle layer of the model to enhance the model's ability to extract key plant features (such as leaf texture and petal shape). The softmax function is used in the output layer to achieve multi-class classification and output species recognition results and confidence scores.

[0121] Details of the improved ResNet-50 model:

[0122] Basic structure: ResNet-50 (5 residual blocks, 177 convolutional layers), input layer 224×224×3 (fused feature maps to 3 channels);

[0123] Improvement details: Input-side multi-scale feature fusion layer: 1×1, 3×3, and 5×5 convolutional kernels extract features in parallel and then concatenate them; Intermediate layer CBAM attention mechanism: Channel attention (global average pooling + 2048→1024→2048 fully connected layer) and spatial attention (7×7 convolution) are added after the residual block; Output layer: Softmax multi-class classification (number of neurons = number of species + 1, including "unidentified" class), confidence score ≥ 0.8 is considered valid recognition, 0.6-0.8 needs to be compared with the sample library similarity (Euclidean distance ≤ 0.5 confirms recognition).

[0124] Model optimization: To address the issue of low accuracy in identifying submerged plants due to dim underwater lighting in wetlands, underwater environment enhancement samples were added to the sample library (using image enhancement algorithms to simulate lighting conditions at different water depths), and an underwater environment adaptive loss function was added during model training to improve the model's adaptability to underwater environments.

[0125] Model training and optimization details:

[0126] Training environment: Ubuntu 22.04 system, PyTorch 2.0 framework, NVIDIA Tesla V100 GPU (32GB VRAM).

[0127] Training parameters: batchsize=32, initial learning rate 0.001 (cosine annealing strategy, decaying to 50% every 10 epochs), epoch=50, cross-entropy loss function, AdamW optimizer (weight decay 0.001).

[0128] Transfer learning: Fine-tuning based on ImageNet pre-trained model, freezing the first 3 residual blocks in the first 10 epochs, and fine-tuning the whole model in the next 40 epochs;

[0129] Validation criteria: Training set:Validation set:Test set = 7:2:1, Test set accuracy ≥ 92% (emergent ≥ 95%, floating ≥ 93%, sinking ≥ 90%).

[0130] Implementation details: Model training uses a GPU server (configured with NVIDIA Tesla V100 graphics cards). Through transfer learning, the model is fine-tuned based on the pre-trained ResNet-50 model to shorten the training cycle. The model is deployed on a backend server and supports two modes: batch processing and real-time recognition. Batch processing is used for species recognition of the full-domain imagery of the multi-rotor UAV1, and real-time recognition is used for fixed-point monitoring data of ground and underwater sensors.

[0131] Model deployment: TensorRT 8.6 acceleration (inference speed improved from 50ms / image to 17ms / image), deployed on an Intel Xeon Gold 6330 server (dual Tesla V100 GPUs), supporting batch processing (1000 images / time) and real-time recognition (≤20ms / image).

[0132] Results verification: 100 samples (30+ each of emergent / floating / sinking) are randomly selected each month for manual verification. If the error rate is >5%, additional samples are added for retraining.

[0133] In the above technical solution: This module relies on a dedicated wetland plant sample library to provide rich and realistic basic data for model training. The samples cover different growth stages and environmental conditions, ensuring that the model can adapt to the complex and ever-changing monitoring scenarios of wetlands. The improved recognition model enhances the ability to extract key plant features by adding a multi-scale feature fusion layer and an attention mechanism, effectively improving the accuracy of species identification. Addressing the difficulty of identification due to low underwater lighting, sample enhancement and model optimization significantly improve the identification effect of submerged plants, solving the pain point of low identification rate for submerged plants in existing technologies. The confidence judgment mechanism of the model output can filter out reliable identification results and mark doubtful results for verification, balancing identification efficiency and accuracy. This module replaces traditional manual species identification work, significantly reducing labor costs and subjective errors, and providing accurate species distribution and quantity information for subsequent diversity index calculations.

[0134] The core indicators of plant diversity calculation module calculates core indicators of wetland plant diversity based on species identification results, including α diversity (species richness, Shannon-Wiener index, Simpson index, Pielou evenness index) and β diversity (Jaccard similarity index, Bray-Curtis distance), comprehensively reflecting the diversity level of wetland plant communities.

[0135] Technical solution:

[0136] Basic data statistics: Based on the species identification results, the number of species and the number of individuals (plants or coverage) of each species in each monitoring plot were counted; a "plot, species" matrix was constructed using the R4.2vegan package (row = plot, column = species, value = number of individuals); emergent / floating plants were counted by plant number, and submerged plants were converted by coverage (1% coverage ≈ 1 plant / 1m²).

[0137] Diversity index calculation: To address the shortcomings of the traditional Shannon-Wiener index, which only considers relative species abundance and ignores the impact of dominant species on the community, this system introduces species dominance weights. The improved Shannon-Wiener index calculation formula is as follows:

[0138] ;

[0139] in, The Shannon-Wiener index is an improved version; a higher value indicates greater species diversity. This represents the total number of species within the sample plot; For the first The relative abundance of each species (the first) (Number of individuals of one species / Total number of individuals of all species in the sample plot); For the first The method for calculating the dominance weight of each species is as follows: For the first Average plant height of each species For the first The coverage of each species. To analyze the species similarity in different habitat areas, the Jaccard similarity index in β-diversity is calculated using the following formula: in, The Jaccard similarity index has a value range of [0,1]. The larger the value, the higher the species similarity between the two plots. The number of species shared by the two plots; The number of species that appear only in the first plot; This represents the number of species appearing only in the second plot. Simultaneously, combining information such as plot area and species distribution, other diversity indicators such as species richness and the Simpson index are calculated. For different wetland areas (e.g., deep water, shallow water, and tidal flats), diversity indicators within each area are calculated to achieve refined zoning analysis.

[0140] Details of core diversity indicator calculation:

[0141] α-diversity index: species richness ( ): Total number of species in the sample plot, to quickly determine the species richness;

[0142] Improved Shannon-Wiener index ( ): ,in (Relative degree) (Dominance weight, Average plant height (coverage) ≥3 indicates high diversity, with values ​​ranging from 0 to 5;

[0143] Simpson index ( ): The higher the value, the less prominent the dominant species are;

[0144] Pielou uniformity index ( ): The range is [0,1], and ≥0.7 indicates uniform distribution.

[0145] β-diversity index: Jaccard similarity index ( ): (a) Number of co-existing species, b) Only present in plot 1, c) Only present in plot 2), ≥0.6 indicates high similarity;

[0146] Bray-Curtis distance ( ): The larger the value, the greater the regional differences.

[0147] Refined partitioning analysis:

[0148] Zoning criteria: Deep water zone (depth > 1.5m, mainly submerged plants); Shallow water zone (0.5-1.5m, floating and submerged plants); Tidal flat zone (<0.5m, mainly emergent plants); ④ Vegetation buffer zone (wetland edge, herbaceous plants and shrubs);

[0149] Analysis process: First calculate the overall indicators, then calculate the indicators for each region, and reveal the distribution characteristics by comparing between regions (such as the BC distance between deep water and shallow water areas) and comparing time series within regions (such as the monthly H' change in tidal flat areas).

[0150] Implementation details: The diversity index was calculated using the R language ecological statistics package (vegan), which automatically associated habitat types in each area with the spatial coordinate information of the sample plots; the calculation results were stored in the database in real time and correlated with historical data to provide data support for subsequent dynamic analysis.

[0151] In the aforementioned technical solution: This module, based on accurate species identification results, conducts refined calculations of diversity indicators. Addressing the deficiency of traditional diversity indices that do not consider wetland plant community characteristics, it introduces dominance weights, making the calculation results more closely reflect the actual community structure of wetland plants and more realistically reflect the roles and positions of different species within the community. In addition to the core indices, multiple diversity indicators are calculated simultaneously to characterize the diversity level of wetland plants from different dimensions. The zonal calculation strategy fully integrates the characteristics of different wetland habitats, accurately presenting the diversity differences in different areas such as deep water, shallow water, and tidal flats, helping managers understand the diversity distribution patterns within wetlands. The indicator data output by this module provides core evidence for subsequent dynamic analysis, shifting the evaluation of wetland plant diversity from qualitative description to quantitative analysis, enhancing the scientific rigor and persuasiveness of monitoring results, and providing crucial data support for wetland ecological status assessment.

[0152] The wetland plant diversity dynamic analysis module analyzes the temporal variation trends (monthly, quarterly, and annual) and spatial distribution characteristics of wetland plant diversity based on historical and real-time monitoring data. It reveals the correlation between diversity changes and environmental factors (water level, soil temperature and humidity, and light), thus solving the problem of the lack of dynamic analysis capabilities in existing technologies.

[0153] Technical solution:

[0154] Time series analysis: Trend fitting and abrupt change test methods were used to analyze the time-varying trends of diversity indicators. Monthly, quarterly, and annual diversity indicator change curves were fitted using linear regression, and the Mann-Kendall test was used to detect abrupt changes in diversity and determine the significance of the changing trends (P<0.05 was considered significant).

[0155] Details of time series analysis:

[0156] Trend Fitting: Linear Regression Formula (k-trend slope, b-intercept), k>0 indicates an upward trend, k<0 indicates a downward trend, |k|≥0.05 indicates a strong trend, 0.02≤|k|<0.05 indicates a medium trend, and |k|<0.02 indicates a weak trend.

[0157] Significance test: Mann-Kendall test (α=0.05, P<0.05 is considered a significant trend), 12-month window analysis of annual trends;

[0158] Mutation point detection: Pettitt test (α=0.05), after detecting mutation points, analyze the corresponding changes in environmental factors (water level, rainfall).

[0159] Spatial distribution analysis: Kriging interpolation was used to interpolate the diversity indices of each monitoring plot into a continuous spatial distribution layer, which visually displays the diversity level of different areas of the wetland; combined with GIS technology, wetland habitat type maps (deep water area, shallow water area, tidal flat area, vegetation buffer zone) were overlaid to analyze the diversity differences under different habitat types.

[0160] Spatial distribution analysis details:

[0161] Interpolation method: ordinary Kriging interpolation, spherical variogram, search radius 50m, generating a 1m×1m raster layer;

[0162] Visualization standards: ≥3 green (high diversity), 2-3 yellow (medium), <2 red (low);

[0163] Spatial pattern analysis: Clustering index ( ), , ≥70 clustered distribution, 50-70 random, <50 uniform.

[0164] Correlation analysis: Redundancy analysis (RDA) was used to analyze the correlation between diversity indicators and environmental factors (water level, soil temperature and humidity, light) to identify key environmental factors affecting wetland plant diversity.

[0165] Correlation analysis details:

[0166] Redundancy Analysis (RDA): The response variable is a diversity indicator ( , , The explanatory variables are environmental factors (water level, soil temperature and humidity, light intensity, rainfall, and water pH).

[0167] Parameter settings: Z-score standardization, 2 ordination axes (explanation rate ≥80%), 999 permutation tests (α=0.05);

[0168] Key factor identification: Key factors are those that explain ≥20% of the variance and have a p-value <0.05 (e.g., water level in the core area and light intensity in the edge area).

[0169] Implementation details: Time series analysis and correlation analysis are implemented using R language, while spatial analysis is implemented using ArcGIS software; the system automatically generates dynamic analysis reports every month, including time trend charts, spatial distribution maps, and key environmental factor analysis results; historical data storage uses a time series database (InfluxDB), supporting efficient time series data querying and analysis.

[0170] Analysis tools: R language vegan / trend package (time series), ArcGIS Pro 3.1 (spatial analysis), Canoco5 (RDA);

[0171] Report generation: Automatically generates dynamic analysis reports every month, including trend charts, spatial distribution maps, and correlation dual-axis plots, and supports PDF export.

[0172] In the aforementioned technical solution: This module conducts dynamic analysis in both time and space dimensions based on historical and real-time monitoring data. Time series analysis clearly reveals the changing trends of wetland plant diversity, accurately identifies key nodes of trend abrupt changes, and helps managers grasp the long-term evolutionary patterns of diversity. Spatial distribution analysis generates continuous distribution layers through interpolation, intuitively presenting the diversity differences in different wetland areas. Combined with the overlay analysis of habitat type maps, it can accurately locate areas with high and low diversity. The correlation analysis stage delves into the intrinsic relationship between diversity indicators and environmental factors, identifies key factors affecting wetland plant diversity, and reveals the interaction patterns between plant growth and environmental change. This module overcomes the limitations of existing static analysis technologies, realizing dynamic monitoring and attribution analysis of diversity, providing a scientific basis for subsequent anomaly warnings and protection recommendations, and making wetland protection work more targeted.

[0173] The diversity anomaly early warning and protection suggestion generation module: Based on dynamic analysis results, it sets diversity anomaly thresholds, monitors abnormal changes in diversity indicators in real time, issues early warning signals, and automatically generates targeted protection suggestions in conjunction with key environmental factor analysis, realizing closed-loop management of "monitoring-early warning-decision", and solving the problem of the lack of early warning and decision support capabilities in existing technologies.

[0174] Technical solution:

[0175] Dynamic trend assessment: Based on the time series analysis results, the trend of diversity changes is determined using the linear regression slope formula, which is as follows:

[0176] ;

[0177] in, The trend slope This indicates an upward trend in diversity. This indicates a downward trend. The larger the value, the more pronounced the trend. To monitor the length of time series (e.g., monthly monitoring), (This represents data from 3 years). For time variables (e.g.) Indicates the first month. (Indicates the second month); For the first The Shannon-Wiener index at each time point.

[0178] Anomaly threshold setting and early warning determination: The mean ± 2 standard deviations of 3 years of historical monitoring data were used as the basic threshold. Combined with wetland ecological protection goals, and revised through expert review, dynamic anomaly thresholds for different regions and seasons were determined. Based on the dynamic thresholds, the following formula was used to determine the early warning level:

[0179] ;

[0180] in, The Shannon-Wiener index is currently being monitored. These are dynamically set thresholds for abnormal data. If an indicator exceeds the threshold range, the system will automatically trigger an alert, which will be sent to administrators via SMS and the mobile app.

[0181] Protection Recommendation Generation: Based on correlation analysis results, key environmental factors leading to biodiversity anomalies are identified. Combined with the ecological habits of wetland plants, a pre-set protection measures knowledge base is invoked to automatically generate targeted protection recommendations (such as adjusting water level control facilities, supplementing irrigation, and planting native plants). Implementation details and anomaly thresholds are updated annually, adjusted based on the year's monitoring data and ecological protection goals. The protection measures knowledge base is established by collecting academic literature, practical cases, and expert experience in the field of wetland ecological protection and is updated regularly. Early warning information is pushed via Alibaba Cloud SMS service and a customized APP to ensure timely receipt by management personnel.

[0182] The monitoring data visualization and interaction module displays species identification results, diversity indicators, dynamic analysis results, and early warning information in an intuitive form. It provides interactive functions such as data query, export, and spatial positioning, making it convenient for managers to view and use monitoring data and improving the system's usability.

[0183] Technical solution:

[0184] Visualization: Utilizing a web-based visualization platform, the displayed content includes:

[0185] Spatial visualization: a species distribution map and a spatial distribution layer of diversity indicators for the entire wetland area, supporting zooming, panning, and click-to-query (clicking on a region will display a list of species and details of diversity indicators for that region);

[0186] Time visualization: Monthly, quarterly, and annual trend charts (line charts and bar charts) for diverse indicators, supporting comparative displays of different indicators and different regions;

[0187] Early warning visualization: Abnormal areas are highlighted and warning level icons are displayed. Clicking on a warning area will display warning details and protection suggestions.

[0188] Statistical Reports: Automatically generates monthly, quarterly, and annual monitoring reports, including data statistics, trend analysis, early warning information, and protection recommendations.

[0189] Interactive functions: Provides data query (by time, region, and species name), data export (supports Excel and PDF formats), spatial positioning (GPS-based plot positioning), and parameter settings (such as adjusting early warning thresholds and setting monitoring frequency).

[0190] The specific details are as follows:

[0191] Platform architecture: a three-tier architecture of "front-end-back-database", with front-end Vue.js 3.2 + ElementPlus, back-end SpringBoot 2.7, and database PostgreSQL + InfluxDB (time series data).

[0192] Page Layout: Homepage (Overview: Diversity Distribution, Real-time Alerts, Key Factors); Species Identification (Distribution Map, Details); Dynamic Analysis (Trends, Spatial Analysis, Relationship Diagram); Alert Management (List, Processing Progress); System Settings (Parameters, Permissions);

[0193] Core Functions: Spatial Visualization: Leaflet 1.9 implements GIS maps, supporting zooming (1:100-1:10000), click query, and switching between species / diversity / environment layers;

[0194] Time visualization: ECharts 5.4 provides line charts / bar charts / pie charts, supports export (PNG / SVG), and allows selection of time ranges from 1 month to 5 years;

[0195] Early warning visualization: Abnormal areas are highlighted with red borders, and early warning icons are distinguished by level (level 1: yellow triangle, level 2: orange circle, level 3: red square), with details displayed in a floating window;

[0196] Interactive functions: multi-condition query (time / region / species), parameter adjustment (early warning coefficient 0.8-1.2, collection frequency 1-4 hours), access control (RBAC: administrator full permissions, operator partial permissions, viewer view only, password MD5 encrypted).

[0197] Implementation details: The visualization platform is developed using the Vue.js framework, combined with ECharts for chart display, and Leaflet for GIS map functionality; the platform is deployed on a cloud server (Alibaba Cloud ECS) and supports access from multiple terminals (computers, tablets, and mobile phones); data export permissions are managed in a hierarchical manner to ensure data security.

[0198] In the aforementioned technical solution: This module transforms complex monitoring data into intuitive and easy-to-understand visualizations. Through spatial distribution maps, time trend charts, and early warning information annotations, it comprehensively displays the monitoring results of wetland plant diversity, enabling managers to quickly grasp the overall monitoring status. The spatial visualization function supports zooming and querying; a single click retrieves detailed information for a specific area, significantly improving data readability. The time visualization function clearly presents the changing trends of diversity indicators, facilitating comparative analysis across different periods and regions. Interactive functions meet the operational needs of different users; data query and export functions facilitate in-depth analysis, and parameter setting functions allow the system to adapt to the monitoring needs of different wetlands. Hierarchical access control ensures data security; different user roles obtain corresponding permissions, avoiding the risk of data leakage. This module enhances the practicality and ease of use of the monitoring system, enabling the efficient utilization of professional monitoring data and providing intuitive support for wetland management decisions.

[0199] System Adaptive Calibration and Maintenance Module: Regularly performs adaptive calibration and maintenance on the system's sensors, recognition models, and analysis algorithms to ensure long-term stable operation of the system, improve the accuracy and reliability of monitoring data, and solve the problem of lack of long-term stability guarantee in existing technologies.

[0200] Technical Solution: Sensor Calibration: On-site calibration of ground and underwater sensors is performed quarterly. Standard equipment (such as standard thermometers and hygrometers, and standard water level gauges) is used to compare sensor measurements. Systematic errors are eliminated using an error correction formula, which is:

[0201] ;

[0202] in, The corrected sensor measurement value. These are the original measurements from the sensor; The sensor error is calculated as follows: The multi-rotor drone 1 and the hyperspectral imager 3 are sent to a professional institution for calibration once a year to ensure the accuracy of the spectral data.

[0203] Model Calibration: New plant samples are collected every six months to supplement the sample bank. The improved CNN model is then retrained and fine-tuned to enhance its ability to identify new species or new growth stages of species. To ensure the reliability of the model's accuracy, a weighted formula is used to calibrate the model's accuracy, combining the original model validation results and the results of manual field verification. The formula is as follows:

[0204] ;

[0205] in, The accuracy of the calibrated model is the recognition rate. This represents the model's original recognition accuracy (based on sample library validation). The accuracy rate of manual on-site verification is calculated by comparing the results of manual identification with those of the model by randomly selecting 100 monitoring samples. The weighting coefficient is 0.7, which is used in this system to balance the basic performance of the model and the objectivity of field validation. The model's recognition accuracy is calculated in real time, and if the accuracy falls below 90%, an emergency calibration process is automatically triggered.

[0206] Algorithm parameter calibration: Based on long-term monitoring data, the weight coefficients of the data fusion algorithm and the parameters of the diversity analysis algorithm are optimized regularly (annually) to make the algorithm more adaptable to changes in the wetland environment; an algorithm error assessment mechanism is established, through manual on-site verification of data and comparison with system monitoring data, errors are calculated and parameters are adjusted.

[0207] System maintenance: Establish an equipment status monitoring mechanism to monitor the operating status of sensors, multi-rotor UAV1, and servers in real time. If equipment failure occurs (such as sensor offline or multi-rotor UAV1 being unable to take off), an automatic maintenance reminder will be issued. The database will be backed up and optimized regularly to ensure data security and system operating efficiency.

[0208] The implementation of detailed sensor calibration adopts standardized operating procedures, and calibration data is recorded and archived; model calibration uses an automated training platform that supports batch data processing and model training; system maintenance adopts a combination of remote monitoring and on-site maintenance. Simple faults (such as sensor offline) can be resolved by remote restart, while complex faults (such as equipment damage) are handled by on-site repair by personnel.

[0209] In the above technical solution: This module establishes a routine calibration and maintenance mechanism to address potential issues such as equipment errors and model performance degradation that may occur during the long-term operation of the monitoring system. Regular sensor calibration promptly corrects equipment deviations, ensuring the accuracy of collected data and extending sensor lifespan. Regular optimization of the identification model and updates to the sample library allow the model to adapt to changes in wetland species, maintaining high identification accuracy and preventing model failure due to species updates or environmental changes. Dynamic optimization of algorithm parameters ensures that data fusion and diversity calculations align with actual wetland changes, guaranteeing the overall analytical accuracy of the system. Real-time equipment status monitoring promptly detects faults and issues alerts, facilitating maintenance work by management personnel and reducing system downtime. This module ensures the long-term stable operation of the monitoring system, guaranteeing the continuous output of accurate and reliable monitoring data, providing a solid technical guarantee for the long-term dynamic monitoring of wetland plant diversity.

[0210] Refer to the experimental setup example:

[0211] I. Overview of the Implementation Scenarios;

[0212] This embodiment selects an inland freshwater lake wetland as the monitoring object. The wetland covers a total area of ​​approximately 120 hm² and is a typical complex habitat wetland, encompassing four habitat types: deep water, shallow water, tidal flats, and a vegetation buffer zone at the wetland edge. The deep water area is dominated by submerged plants such as *Myriophyllum spicatum* and *Ceratophyllum demersum*, while the shallow water area is home to floating plants such as lotus and *Euryale ferox*. The tidal flats are covered by emergent plants such as reeds and cattails, and the buffer zone is dominated by a mixed community of herbaceous shrubs. The wetland is surrounded by agricultural planting areas, posing potential disturbances such as agricultural non-point source pollution and water level fluctuations. Accurate monitoring of plant diversity changes is necessary to support ecological protection.

[0213] II. System Deployment Details;

[0214] Multi-source data collaborative acquisition unit deployment: A three-in-one architecture of "air, ground, and underwater" is adopted. The air acquisition unit uses a multi-rotor UAV1 equipped with a hyperspectral imager3 and an RGB high-definition camera2, planning a combination of "circular and radial" flight paths based on the irregular lake morphology of the wetland to cover the entire area without blind spots; the ground acquisition unit is deployed according to the "grid method and habitat stratification," with denser deployment in the core area and conventional deployment in the edge area. Each plot is equipped with soil temperature and humidity sensors, light sensors, and water level sensors, fixed with stainless steel brackets to ensure adaptation to the wetland's humid environment; the underwater acquisition unit is deployed according to the water depth gradient in deep water areas, shallow water areas, and transition zones, with underwater camera lenses pointing upwards to capture the canopy of submerged plants, and spectral sensors fixed to avoid areas of siltation; the data transmission unit adopts "5G+LoRa" dual-mode, deploying LoRa gateways around the wetland to ensure full signal coverage, and all collected data is encrypted before being transmitted to the backend server.

[0215] Other module deployments: The data preprocessing and multi-source data fusion module is deployed at the edge gateway to process raw data locally and reduce transmission pressure; the wetland plant species intelligent identification module is deployed on the back-end GPU server, equipped with a trained improved ResNet-50 model, which has been fine-tuned based on 32 native plant samples of the wetland; the plant diversity core indicator calculation module, dynamic analysis module, and anomaly warning and protection suggestion generation module are integrated in the back-end data processing center and linked with the visualization interaction platform; the system adaptive calibration and maintenance module is deployed synchronously in the back-end, with preset calibration cycles and maintenance processes to ensure long-term stable operation of the system.

[0216] III. System Operation Flow;

[0217] Data collection phase: Based on the wetland's climate characteristics, comprehensive data collection is conducted twice a month during the non-flood season (April-May and September-October). During the flood season (June-August), due to large water level fluctuations, data is collected once a month. In winter, data is collected once every two months. The collection time is fixed at 10:00-14:00 daily during periods of stable sunlight. A multi-rotor UAV 1 completes aerial image collection according to a preset route, while ground and underwater sensors continuously collect environmental parameters and plant growth status data at set frequencies. All data is transmitted back in real time with encryption.

[0218] Data processing and analysis phase: The edge gateway preprocesses the returned raw data to eliminate atmospheric scattering, water refraction, and sensor noise interference. After calculating the weights of each data source using the entropy weight method, multi-source data fusion is completed to generate a standardized dataset. The standardized dataset is transmitted to the backend server, where an improved ResNet-50 model is used to identify species, outputting information such as species name, distribution location, and growth status. Based on the identification results, core diversity indicators such as the Shannon-Wiener index with dominance weights are calculated. Simultaneously, time-series analysis, spatial distribution analysis, and correlation analysis are conducted to accurately characterize diversity change trends and spatial distribution patterns, and identify key environmental factors affecting the wetland's plant diversity.

[0219] Early warning and decision-making phase: The system constructs dynamic thresholds based on three years of historical monitoring data, and completes expert review and correction in conjunction with the ecological protection goals of the wetland, while monitoring the deviation between diversity indicators and thresholds in real time. When a moderate anomaly is detected in the diversity indicators of floating plants in shallow water areas, the system automatically triggers a level-two early warning, notifying wetland management personnel via SMS, APP push notifications, etc. Based on the key influencing factors (water level fluctuations) identified by correlation analysis, the system generates targeted protection recommendations from the protection measures knowledge base, including specific implementation steps and expected effects such as adjusting surrounding water conservancy control facilities and stabilizing the water level in shallow water areas.

[0220] Calibration and maintenance phase: System maintenance is carried out according to the preset cycle. Ground and underwater sensors are calibrated on-site every quarter. The multi-rotor UAV 1 and hyperspectral imager 3 are sent to a professional institution for calibration every year. New plant samples in the wetland are added every six months to retrain and fine-tune the recognition model. The back-end server is inspected weekly and the database is backed up daily to ensure long-term stable operation of the system.

[0221] IV. Implementation Results;

[0222] The implementation of this system in this freshwater lake wetland has enabled comprehensive, all-encompassing monitoring of emergent, floating, and submerged plants, completely solving the problems of traditional manual monitoring's inability to cover deep water areas and its low efficiency. The accuracy of species identification aligns with the actual habitat requirements of the wetland, effectively distinguishing morphologically similar wetland plant species. Dynamic analysis and a tiered early warning mechanism enable early detection and rapid response to diversity anomalies, preventing the escalation of ecological problems. The targeted and practical protection recommendations provide clear action guidelines for wetland managers, significantly improving the scientific rigor and precision of the wetland's ecological protection efforts.

[0223] In summary, this invention fills the gap in submerged plant monitoring through a multi-source data collaborative acquisition module, achieving comprehensive, blind-spot-free monitoring of emergent, floating, and submerged plants. Dual-mode communication balances efficient transmission with low power consumption, minimizing human intervention and interference with the wetland ecosystem. The collected multi-dimensional data lays a solid foundation for subsequent analysis. The data preprocessing and multi-source data fusion module specifically eliminates environmental interference such as atmospheric scattering and water refraction, as well as equipment errors. Differential optimization strategies enhance the distinction between plants and the background, and the objective entropy weight method organically integrates multi-source data to generate a standardized fusion dataset. This overcomes the limitations of single data sources, adds value to the data, and provides high-quality input for subsequent analysis. The intelligent wetland plant species identification module, with its dedicated wetland plant sample library, adapts to complex monitoring scenarios. Improved model structure enhances key feature extraction, and sample augmentation technology addresses the pain points of submerged plant identification. The confidence judgment mechanism balances efficiency and accuracy. The core plant diversity indicator calculation module improves upon traditional indices by introducing dominance weights to align with the realities of wetland communities. It employs refined zoning calculations to reveal differences in diversity across different habitats, achieving a scientific evaluation from qualitative to quantitative perspectives and providing a core basis for dynamic analysis. The wetland plant diversity dynamic analysis module conducts temporal and spatial analysis, accurately revealing trends, identifying key nodes and regional differences, and deeply exploring the intrinsic relationship between diversity and environmental factors. This overcomes the limitations of static analysis and provides a scientific basis for early warning and decision-making. The diversity anomaly early warning and protection suggestion generation module uses dynamic thresholds to adapt to environmental changes, a tiered early warning mechanism to accurately convey the degree of anomaly, and multi-channel push notifications to ensure timely response. The monitoring data visualization and interaction module transforms complex data into intuitive visualizations of spatial distribution and temporal trends. It supports multi-condition queries, data export, and parameter settings, and combines RBAC permission management to ensure data security, significantly improving the system's practicality and ease of use, allowing professional data to efficiently serve management decisions. The system adaptive calibration and maintenance module establishes a routine calibration and maintenance mechanism, regularly correcting sensor errors, optimizing recognition models and algorithm parameters, monitoring equipment status in real time, promptly troubleshooting, ensuring long-term stable system operation, and guaranteeing continuously accurate and reliable monitoring data.

[0224] Although embodiments of the invention have been shown and described, the scope of the invention will be defined by the appended claims and their equivalents by those skilled in the art.

Claims

1. A wetland plant diversity monitoring device, characterized in that, It includes a multi-source data acquisition unit, a data transmission unit, and a data encryption unit; The multi-source data acquisition unit includes an aerial acquisition unit, a ground acquisition unit, and an underwater acquisition unit; The aerial acquisition unit includes a multi-rotor drone (1), and a hyperspectral imager (3) and an RGB high-definition camera (2) mounted on the multi-rotor drone (1). The ground acquisition unit includes a sensor array and a portable hyperspectral analyzer deployed in typical wetland sample plots. The sensor array includes a soil temperature and humidity sensor, a light sensor, and a water level sensor. The underwater acquisition unit includes underwater high-definition cameras and underwater spectral sensors deployed at different water depth gradients in the wetland. The data transmission unit adopts a 5G and LoRa dual-mode communication architecture to transmit the data collected by the multi-source data acquisition unit to the backend server. The data encryption unit uses the AES-128 encryption algorithm to encrypt data before transmission.

2. The wetland plant diversity monitoring system according to claim 1, characterized in that, It includes the wetland plant diversity monitoring device as described in claim 1, a data preprocessing and multi-source data fusion module, a wetland plant species intelligent identification module, a plant diversity core indicator calculation module, a wetland plant diversity dynamic analysis module, a diversity anomaly early warning and protection suggestion generation module, a monitoring data visualization and interaction module, and a system adaptive calibration and maintenance module. The data preprocessing and multi-source data fusion module is used to perform noise reduction, correction and registration preprocessing on the original data, calculate the weight of each data source using the entropy weight method, and then complete the multi-source data fusion based on the weighted fusion formula to generate a standardized dataset. The intelligent wetland plant species identification module is based on an improved ResNet-50 model to identify emergent, floating, and submerged plants in wetlands. The core plant diversity index calculation module is used to calculate the Shannon-Wiener index and Jaccard similarity index with species dominance weights. The wetland plant diversity dynamic analysis module is used to analyze the temporal variation trend and spatial distribution characteristics of diversity indicators; The diversity anomaly early warning and protection suggestion generation module realizes hierarchical early warning based on trend slope calculation results and dynamic thresholds, and automatically generates protection suggestions; The monitoring data visualization and interaction module is used to realize spatial visualization, temporal visualization and early warning visualization of monitoring data; The system's adaptive calibration and maintenance module is used to correct sensor errors and calibrate the accuracy of the recognition model.

3. The wetland plant diversity monitoring system according to claim 2, characterized in that, The specific processing procedure of the data preprocessing and multi-source data fusion module is as follows: Data preprocessing: Radiometric correction, atmospheric correction and geometric correction were performed on the hyperspectral images of the multi-rotor UAV (1); noise reduction and brightness equalization were performed on the RGB images and underwater images; Sensor data preprocessing: Outlier removal and missing value completion are performed on time-series data such as soil temperature and humidity, light intensity, and water level, and the data is then standardized. After multi-source data fusion preprocessing, hyperspectral data, RGB morphological data, and environmental sensor data need to be weighted and fused. The entropy weight method is used to calculate the weight of each data source. The weight calculation formula is as follows: ； in, For the first The weight coefficients of each data source must satisfy... For the first The information entropy of each data source is calculated as follows: The number of data sources to be merged; For the first The number of samples from each data source; For the first The first data source Normalized frequencies of each sample; Based on the theory of information entropy, the lower the information entropy of a data source, the lower its uncertainty and the higher its effective information content, and the larger its corresponding weight coefficient. After determining the weights, the following formula is used to complete the fusion of multi-source data and generate a standardized dataset containing three-dimensional information: ； in, For the merged standardized data, For the first Standardized data from a data source after preprocessing.

4. The wetland plant diversity monitoring system according to claim 3, characterized in that, The specific process of the intelligent identification module for wetland plant species is as follows: Establish a wetland plant sample bank: Collect hyperspectral data, morphological data and growth environment data of common plant species in the target wetland, and construct an labeled sample bank with samples covering different growth stages and different environmental conditions; Improved CNN recognition model: An improved ResNet-50 model is used as the base network. A multi-scale feature fusion layer is added at the model input to fuse the spectral features of hyperspectral data and the morphological features of RGB data. An attention mechanism is introduced in the middle layer of the model to enhance the model's ability to extract key plant features. The softmax function is used in the output layer to achieve multi-class classification and output species recognition results and confidence scores. Model optimization: To address the issue of low accuracy in identifying submerged plants due to dim underwater lighting in wetlands, underwater environment enhancement samples were added to the sample library, and an underwater environment adaptive loss function was incorporated into the model training process.

5. A wetland plant diversity monitoring system according to claim 4, characterized in that, The specific process of the core plant diversity index calculation module is as follows: Basic data statistics: Based on the species identification results, the number of species and the number of individuals of each species in each monitoring plot are counted; Diversity index calculation: To address the shortcomings of the traditional Shannon-Wiener index, which only considers relative species abundance and ignores the impact of dominant species on the community, a species dominance weight is introduced. The improved Shannon-Wiener index calculation formula is as follows: ； in, The Shannon-Wiener index is an improved version; a higher value indicates greater species diversity. This represents the total number of species within the sample plot; For the first The relative abundance of each species; For the first The dominance weights of each species are calculated as follows: For the first Average plant height of each species For the first Coverage of individual species; To analyze species similarity across different habitat regions, the Jaccard similarity index in β-diversity is calculated using the following formula: in, The Jaccard similarity index has a value range of [0,1]. The larger the value, the higher the species similarity between the two plots. The number of species shared by the two plots; This represents the number of species that appear only in the first plot; This represents the number of species that appeared only in the second plot; By combining information such as plot area and species distribution, we can calculate species richness, Simpson index and other diversity indicators. For the habitat differences in different wetland areas, we can calculate the diversity indicators in each area to achieve zoning analysis.

6. A wetland plant diversity monitoring system according to claim 5, characterized in that, The specific process of the wetland plant diversity dynamic analysis module is as follows: Time series analysis: Trend fitting and abrupt change test methods are used to analyze the time-varying trend of diversity indicators; monthly, quarterly and annual diversity indicator change curves are fitted by linear regression, and the Mann-Kendall test method is used to detect abrupt change points in diversity changes and determine the significance of the change trend. Spatial distribution analysis: Kriging interpolation was used to interpolate the diversity indicators of each monitoring plot into a continuous spatial distribution layer, which visually shows the diversity level of different areas of the wetland; combined with GIS technology, wetland habitat type maps were overlaid to analyze the diversity differences under different habitat types. Correlation analysis: Redundancy analysis was used to analyze the correlation between diversity indicators and environmental factors, and to identify key environmental factors affecting wetland plant diversity.

7. A wetland plant diversity monitoring system according to claim 6, characterized in that, The specific process of the diversity anomaly early warning and protection suggestion generation module is as follows: Dynamic trend assessment: Based on the time series analysis results, the trend of diversity changes is determined using the linear regression slope formula, which is as follows: ； in, The trend slope This indicates an upward trend in diversity. This indicates a downward trend. The larger the value, the more pronounced the trend. To monitor the length of the time series; It is a time variable; For the first Shannon-Wiener index at each time point; Anomaly threshold setting and early warning determination: The mean ± 2 standard deviation of 3 years of historical monitoring data is used as the basic threshold. Combined with wetland ecological protection goals, and revised through expert review, dynamic anomaly thresholds for different regions and seasons are determined. Based on the dynamic thresholds, the following formula is used to determine the early warning level: ； in, The Shannon-Wiener index is currently being monitored. The abnormal threshold is dynamically set; if the indicator exceeds the threshold range, the system will automatically trigger an alert, and the alert information will be pushed to the management personnel via SMS and APP. Protection Recommendation Generation: Based on the results of correlation analysis, key environmental factors leading to abnormal diversity are identified. Combined with the ecological habits of wetland plants, a pre-set knowledge base of protection measures is invoked to automatically generate targeted protection recommendations.

8. A wetland plant diversity monitoring system according to claim 7, characterized in that, The specific process of the monitoring data visualization and interaction module is as follows: Visualization: Utilizing a web-based visualization platform, the displayed content includes: Spatial visualization: a species distribution map of the entire wetland area and a spatial distribution layer of diversity indicators, supporting zooming, panning, and click-to-query; Time visualization: Monthly, quarterly, and annual trend charts for diverse indicators, supporting comparative displays of different indicators and regions; Early warning visualization: Abnormal areas are highlighted and warning level icons are displayed. Clicking on a warning area will display warning details and protection suggestions. Statistical Reports: Automatically generates monthly, quarterly, and annual monitoring reports, including data statistics, trend analysis, early warning information, and protection recommendations; Interactive functions: Provides data query, data export, spatial positioning, and parameter setting functions.

9. A wetland plant diversity monitoring system according to claim 8, characterized in that, The specific process of the system's adaptive calibration and maintenance module is as follows: Sensor calibration: On-site calibration of ground and underwater sensors is performed quarterly. Sensor measurements are compared using standard equipment, and systematic errors are eliminated using an error correction formula: ； in, The corrected sensor measurement value. These are the original measurements from the sensor; The sensor error is calculated as follows: ; Model calibration: New plant samples are collected every six months to supplement the sample bank. The improved CNN model is then retrained and fine-tuned to enhance its ability to identify new species or new growth stages of species. Combining the original model validation results with the results of manual field verification, a weighted formula is used to calibrate the model's accuracy. The formula is as follows: ； in, The accuracy of the calibrated model is the recognition rate. This represents the model's original recognition accuracy. To improve the accuracy of manual on-site verification, The weighting coefficients are used to calculate the model's recognition accuracy in real time. If the accuracy falls below 90%, an emergency calibration process is automatically triggered. Algorithm parameter calibration: Based on long-term monitoring data, the weight coefficients of the data fusion algorithm and the parameters of the diversity analysis algorithm are regularly optimized to make the algorithm more adaptable to changes in the wetland environment; an algorithm error assessment mechanism is established, through manual on-site verification of data and comparison with system monitoring data, errors are calculated and parameters are adjusted; System maintenance: Establish an equipment status monitoring mechanism to monitor the operating status of sensors, multi-rotor UAVs (1), and servers in real time. If equipment failure occurs, an automatic maintenance reminder will be issued. The database will be backed up and optimized regularly.

Images