Wetland environment condition acquisition, investigation and analysis method
By processing UAV imagery data and using machine learning models, a dynamic early warning mechanism for wetland water quality monitoring was established, which solved the problem of poor timeliness in wetland water quality monitoring and enabled real-time identification and efficient management of water quality anomalies.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-04
- Publication Date
- 2026-03-13
AI Technical Summary
Current technologies for wetland water quality monitoring rely on manual sampling and laboratory analysis, lacking a dynamic early warning mechanism based on optical characteristics. This results in poor timeliness and insufficient spatial coverage for detecting water quality anomalies, making it difficult to meet the real-time monitoring and management needs of large-scale wetland ecosystems.
Wetland image data is collected by drones, and color-fidelity images are generated through aerial triangulation and orthophoto mosaicking. A machine learning model is then used to establish an inversion model between the RGB optical characteristics of water bodies and water quality indicators, thereby generating spatial distribution maps of water quality indicators and outputting early warning signals.
It enables real-time identification and early warning of areas with abnormal water quality, improves monitoring efficiency, reduces manpower and time costs, and provides an efficient means of wetland water environment management.
Smart Images
Figure CN121656243A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wetland water environment investigation and analysis technology, and in particular to methods for collecting, investigating and analyzing wetland environmental conditions. Background Technology
[0002] As the vital role of wetland ecosystems in water environment regulation and biodiversity conservation becomes increasingly prominent, dynamic monitoring and refined management of wetland water quality have become important directions in ecological and environmental research. While traditional water quality monitoring methods can provide high-precision location data, they are insufficient to meet the demands of the complex environment of wetlands, which exhibits strong spatial heterogeneity and high frequency of change. In recent years, unmanned aerial vehicle (UAV) remote sensing technology has gradually become an important tool for water quality monitoring due to its advantages such as high resolution, flexibility, and real-time performance. Combining machine learning algorithms with the analysis of hyperspectral or visible light images acquired by UAVs enables quantitative inversion between water body optical characteristics and water quality indicators, providing a new technical approach and application foundation for constructing an efficient and automated wetland water quality monitoring system.
[0003] In existing technologies, water quality monitoring mainly relies on manual sampling and laboratory analysis, or indirect estimation using traditional remote sensing images. It lacks a dynamic early warning mechanism based on optical characteristics, resulting in problems such as poor timeliness, insufficient spatial coverage, and delayed response in water quality anomaly detection. It is difficult to detect potential pollution or eutrophication risks in wetland water bodies in a timely manner, and the monitoring results mostly rely on manual interpretation with low automation, making it difficult to meet the real-time monitoring and management needs of large-scale wetland ecological environments. Summary of the Invention
[0004] To overcome the above shortcomings, this invention provides a method for collecting, investigating and analyzing wetland environmental conditions, aiming to improve the problem that existing technologies for water quality monitoring mainly rely on manual sampling and laboratory analysis, and lack a dynamic early warning mechanism based on optical characteristics.
[0005] This invention provides the following technical solution: a method for collecting, investigating, and analyzing wetland environmental conditions, including: S1. Collect UAV imagery data of the wetland study area, and collect in-situ water quality index data and geographic coordinates of multiple sampling points within the wetland; S2. Perform aerial triangulation and orthomorphic mosaicking on the UAV image data to generate an orthomorphic mosaic image. S3. Based on the color calibration plate reference area in the orthophoto mosaic image, perform white balance correction on the orthophoto mosaic image to obtain a color-accurate orthophoto mosaic image. S4. Based on the geographic coordinates of the sampling points, extract the RGB color values of the corresponding pixels from the color-fidelity orthophoto mosaic image, and associate and match the RGB color values with the in-situ water quality index data to construct a water quality inversion training dataset. S5. Use the water quality inversion training dataset to train a machine learning model and establish an inversion model between the RGB optical features of water bodies and water quality indicators. S6. Apply the inversion model to the color-fidelity orthophoto mosaic image to generate a spatial distribution map of water quality indicators, and output a warning signal based on the comparison results of the inversion indicator values in the spatial distribution map and the optical warning threshold.
[0006] Preferably, in step S2, performing aerial triangulation and orthophoto mosaicking on the UAV image data to generate an orthophoto mosaic image specifically includes the following steps: The drone image data is subjected to quality screening to remove abnormal images; Import the filtered image data, generate a sparse point cloud model through feature matching, and perform bundle adjustment of the regional network. The sparse point cloud model and camera pose parameters are used to perform 3D reconstruction, generating a dense point cloud and texture model. The images are geometrically corrected and color balanced, and overlapping areas are joined together to create an orthophoto mosaic.
[0007] Preferably, in step S3, performing white balance correction on the orthophoto based on the color calibration plate reference area in the orthophoto to obtain a color-accurate orthophoto specifically includes the following steps: Identify and locate the color calibration plate reference area in the orthophoto mosaic image; The average RGB color value is extracted from the grayscale block of the reference area of the color calibration plate as the image measurement value; The RGB color values of the grayscale block are obtained as theoretical reference values; Based on the difference between the image measurement value and the theoretical reference value, the white balance gain coefficients of the R, G, and B channels are calculated. The white balance gain coefficient is applied to the orthophoto mosaic image to obtain a color-fidelity orthophoto mosaic image.
[0008] Preferably, in step S4, the specific steps for extracting the RGB color values of corresponding pixels from the color-fidelity orthophoto mosaic image based on the geographic coordinates of the sampling points include: The geographic coordinates of the sampling points are mapped onto the color-fidelity orthophoto mosaic image through coordinate transformation to determine the corresponding center pixel coordinates. Using the coordinates of the center pixel as a reference, a pixel window region of a preset size is defined; Extract the RGB values of all pixels within the pixel window region; Calculate the average RGB value of all pixels within the pixel window area, and determine the average RGB value as the RGB color value of the corresponding pixel.
[0009] Preferably, in step S4, the specific steps for associating and matching the RGB color values with the in-situ water quality index data to construct a water quality inversion training dataset include: For each of the sampling points, create a data record; The RGB color value corresponding to the sampling point is stored in the data record as an input feature vector; The in-situ water quality index data corresponding to the sampling point are stored in the data record as the expected output label; Data records from all the sampling points are collected to generate a water quality inversion training dataset.
[0010] Preferably, in step S5, the specific steps for training a machine learning model using the water quality inversion training dataset to establish an inversion model between the RGB optical features of water bodies and water quality indicators include: The water quality inversion training dataset is divided into a training subset and a validation subset; A machine learning model is selected, and the training subset is input into the machine learning model for training; The model in training is evaluated using the validation subset, and the model hyperparameters are adjusted based on the evaluation results. When the performance metrics of the machine learning model meet the convergence condition, training is stopped, and the model is identified as an inversion model.
[0011] Preferably, in step S6, the specific steps of applying the inversion model to the color-fidelity orthophoto mosaic image to generate a spatial distribution map of water quality indicators include: Traverse the pixels in the color-fidelity orthophoto mosaic image; Extract the RGB color values of the pixels and use them as input features; The input features are input into the inversion model to obtain the inverted water quality index value corresponding to the pixel; The retrieved water quality index values corresponding to all pixels are collected and mapped to the corresponding geographic coordinates to generate a spatial distribution map of the water quality index.
[0012] Preferably, the specific steps for outputting a warning signal based on the comparison result between the inversion index value in the spatial distribution map and the optical warning threshold include: The inverted water quality index values in the spatial distribution map are compared one by one with the optical early warning threshold; Identify and mark areas where the retrieved water quality index values exceed the optical warning threshold; When the marked area is detected, a warning signal is generated and output.
[0013] The present invention has the following beneficial effects: 1. In this invention, by automatically comparing the spatial distribution map of water quality generated by inversion with the preset optical early warning threshold, the real-time identification, spatial positioning and early warning signal output of water quality abnormal areas are realized. It can proactively discover potential environmental risks such as eutrophication and abnormal turbidity in the early stage, and provide intuitive and operable decision support for wetland ecological protection and pollution prevention and control.
[0014] 2. In this invention, by implementing rigorous aerial triangulation, orthorectification, and white balance color calibration, the geometric distortion and spectral distortion of the image are effectively eliminated, ensuring that the RGB optical features used for model training have high spatial consistency and color fidelity, thereby significantly enhancing the prediction accuracy, stability, and generalization ability of the water quality inversion model to different wetland environments and lighting conditions.
[0015] 3. In this invention, by constructing a machine learning inversion model based on UAV color fidelity images and measured water quality parameters, a precise quantitative relationship between the RGB optical characteristics of water bodies and multiple water quality indicators is established, realizing rapid and non-destructive water quality monitoring of large-scale water areas. This completely changes the traditional lagging mode that relies on manual sampling and laboratory analysis, significantly improves monitoring efficiency, reduces manpower and time costs, and provides a highly efficient and scalable technical means for wetland water environment management. Attached Figure Description
[0016] Figure 1 This is a flowchart of the wetland environmental data collection, investigation, and analysis method proposed in this invention. Detailed Implementation
[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] In the first embodiment of the present invention, the present invention provides a method for collecting, investigating and analyzing wetland environmental conditions, such as... Figure 1 As shown, it includes the following steps: S1. Collect UAV imagery data of the wetland study area, and collect in-situ water quality index data and geographic coordinates of multiple sampling points within the wetland; Specifically, firstly, a typical wetland study area was selected as the experimental area. Based on the topographic features, water distribution, and vegetation cover of the study area, a UAV aerial survey plan was developed, determining parameters such as flight altitude, flight path planning, image overlap rate, and ground resolution. To ensure the geometric accuracy and color fidelity of the images, the flight altitude was set to 80 to 120 meters, the forward overlap rate to 80%, and the lateral overlap rate to 70%. The UAV was equipped with a high-resolution visible light camera, acquiring images in RAW or high-quality JPEG format to fully record the true spectral information of the water surface layer.
[0019] Before aerial surveying, the camera needs to be radiometrically calibrated and its white balance set to eliminate color shifts caused by differences in ambient lighting. Simultaneously, several ground control points are established, and their geographic coordinates are acquired using high-precision GNSS positioning equipment to improve the spatial accuracy of subsequent orthophotos. The UAV automatically executes the aerial survey mission according to a preset route, acquiring multiple highly overlapping image datasets covering the entire wetland study area, providing raw image data for generating color-accurate orthophoto mosaics.
[0020] Multiple in-situ water quality sampling points were deployed within the wetland area simultaneously with or around the time of drone image acquisition. Sampling points should cover areas with different water body types and varying optical characteristics, such as open water surfaces, shallow areas, and vegetated areas. A multi-parameter water quality analyzer was used for on-site measurements, recording water quality indicators such as chlorophyll a concentration, turbidity, suspended solids concentration, conductivity, dissolved oxygen, and water temperature at each sampling point, along with simultaneous measurement of their geographic coordinates. To ensure the correspondence between drone image reflection information and water quality measurement data, sampling should be conducted during sunny, stable lighting periods, with minimal disturbance to the water surface.
[0021] By matching geographic coordinates, the measured water quality data and UAV imagery are spatially registered to form a basic water quality inversion dataset. This dataset can be represented as: ; in, : Represents the basic dataset for water quality inversion; : Indicates the total number of sampling points; : indicates the first The RGB optical feature vector of each pixel corresponding to a sampled image point, where , , These are the reflection components for the red, green, and blue bands, respectively. : indicates the first The measured water quality index vector of each sampling point, where This refers to the concentration of chlorophyll a. Turbidity The concentration of suspended solids. Electrical conductivity; : indicates the first The geographical coordinates of each sampling point.
[0022] By combining UAV aerial surveys with simultaneous ground sampling, high-resolution, high-overlap visible light images covering the entire wetland study area were acquired, along with various in-situ water quality parameters (such as chlorophyll a and turbidity) precisely corresponding to their spatial locations. This formed a geospatially rigorous, multimodal foundational dataset. This provides an indispensable and highly spatiotemporally consistent data foundation for subsequently establishing a quantitative relationship model between "image optical features and water quality parameters," fundamentally ensuring the reliability and accuracy of the inversion model.
[0023] S2. Perform aerial triangulation and orthomorphic mosaic processing on the UAV image data to generate orthomorphic mosaic images. Furthermore, in S2, the aerial triangulation and orthomorphic mosaicking of the UAV image data to generate an orthomorphic mosaic image specifically includes the following steps: Perform quality screening on drone imagery data and remove abnormal images; Import the filtered image data, generate a sparse point cloud model through feature matching, and perform bundle adjustment of the regional network. 3D reconstruction is performed using sparse point cloud models and camera pose parameters to generate dense point clouds and texture models. The images are geometrically corrected and color balanced, and overlapping areas are joined together to create an orthophoto mosaic.
[0024] Specifically, in this invention, step S2 first involves rigorous quality screening of the original image sequence acquired by the UAV. The system automatically or manually removes images with motion blur, abnormal exposure, or cloud obstruction using image sharpness assessment and exposure analysis algorithms, ensuring that the data processed subsequently has good geometric and radiometric quality. The screened images are then imported into a photogrammetric processing system, which uses algorithms such as Scale Invariant Feature Transform (SFT) or Orb to extract stable feature points from the images and utilizes fast matching algorithms such as... Tree or FLANN is used to match feature points in images with overlapping regions, thus establishing a sparse point cloud model. Then, the Bundle Adjustment algorithm is executed to perform nonlinear joint optimization on the matching results, improving the geometric accuracy of the image. This algorithm optimizes the 3D point cloud model by minimizing the reprojection error of all feature points. coordinates and camera extrinsic parameters The objective function is: ; in, Two-dimensional coordinates of feature points in the image For three-dimensional points In camera parameters Projected coordinates below The spatial coordinate vector of the three-dimensional feature points in the scene. For camera external parameters, A robust kernel function to reduce the interference of outliers. This represents the total number of matching points participating in the adjustment. Through this optimization process, the system can obtain pose parameters with higher accuracy than the original GPS and MU records from the UAV, providing high-quality geometric constraints for subsequent 3D reconstruction.
[0025] After sparse point cloud optimization, the system enters the dense matching stage. This process employs semi-global matching (SGM) or patch-based multi-view stereo vision (PMS) algorithms to perform pixel-level depth estimation on the image, generating a high-density 3D point cloud. Furthermore, a digital surface model (DSM) or a textured triangular mesh model (Msh) is constructed for surface reconstruction and orthorectification. In the orthorectification stage, geometric transformations are performed based on the collinearity equation principle to ensure the spatial accuracy of image pixel positions. The geometric mapping model can be represented as: ; in, These are the two-dimensional coordinates of the pixels in the original image. These are the pixel coordinates on the orthophoto plane. For camera internal and external parameters, For digital surface models, This is a spatial transformation function from image coordinates to geographic coordinates. Through this process, the system effectively eliminates geometric distortions caused by camera tilt and terrain undulations, generating geometrically accurate orthorectified images.
[0026] To eliminate brightness and tone differences caused by different flight paths or shooting times, the system performs color equalization and illumination homogenization on the images. This process establishes a linear radiometric correction model based on the radiometric characteristics of overlapping areas. ; in, This is the vector of pixel values for the red, green, and blue channels of the original image. These are the corrected pixel values. The image gain coefficient matrix is used to adjust the brightness ratio. This is a bias vector used to correct the overall tone shift. Coefficients and The system obtains the solution from pixel pairs in the overlapping area of the images using the least squares method, thereby achieving brightness uniformity and color balance among multiple images. After color correction, the system calculates the optimal mosaic line within the overlapping area and uses Dijkstra's shortest path algorithm to search for the minimum cost path in the pixel grid. The cost function takes into account factors such as color difference, texture gradient, and smoothness to ensure visual continuity of the stitched area.
[0027] After the mosaic lines are determined, the system applies feathering blending techniques on both sides of the mosaic lines to achieve a smooth transition, such as multi-resolution spline fusion or alpha blending, to eliminate seams and color banding. Finally, all images that have undergone geometric correction, color equalization, and mosaicking are merged according to their precise geographic coordinates to generate a complete orthophoto mosaic image. The output is in GeoTIFF format, containing high-precision RGB information for each pixel and geospatial reference data, such as WGS84 coordinate system and UTM projection. This orthophoto mosaic image ensures a one-to-one correspondence between each pixel and the actual geographic location of the wetland area, providing a unified, reliable, and spatially continuous data foundation for the subsequent white balance correction in step S3 and the extraction of RGB values based on sampling points and the construction of a water quality inversion training dataset in step S4.
[0028] S3. Based on the color calibration plate reference area in the orthophoto mosaic image, perform white balance correction on the orthophoto mosaic image to obtain a color-accurate orthophoto mosaic image. Furthermore, in S3, based on the color calibration plate reference area in the orthophoto mosaic image, white balance correction is performed on the orthophoto mosaic image to obtain a color-accurate orthophoto mosaic image. This specifically includes the following steps: Identify and locate the color calibration plate reference area in the orthophoto mosaic image; The average RGB color value is extracted from the grayscale block of the color calibration panel reference area as the image measurement value; Obtain the RGB color values of the grayscale block as the theoretical baseline value; Based on the difference between the image measurement values and the theoretical reference values, the white balance gain coefficients of the R, G, and B channels are calculated. By applying the white balance gain coefficient to the orthophoto, a color-accurate orthophoto is obtained.
[0029] Specifically, using the high-precision orthophoto mosaic image generated in step S2 as input data, white balance correction is performed on the orthophoto mosaic image based on the color calibration plate reference area set in the image to obtain a color-accurate orthophoto mosaic image. By extracting the color information of the standard gray block in the calibration plate, a mapping relationship between the actual observed color and the theoretical standard color is established, thereby correcting the color cast problem caused by lighting conditions, camera response differences or post-processing between different images, and ensuring the consistency and authenticity of the image colors.
[0030] First, the color calibration plate reference area is automatically identified and located in the orthophoto mosaic image. Since the calibration plate is usually composed of high-contrast rectangular grayscale blocks, a detection algorithm based on color and shape constraints is used to identify feature regions in the image and accurately locate the position of each grayscale block. Each identified grayscale block region is denoted as... The system calculates the average RGB value of all pixels within the grayscale block as the image measurement value of that grayscale block, expressed as: ; in, For the first The number of pixels in a grayscale block For the first The first grayscale block 1 pixel vector, For the first The average observed color value of each grayscale block is used, and simultaneously, the theoretical reference color value of the corresponding grayscale block is read from the standard color calibration chart model, denoted as . This value reflects the ideal reflectivity under standard lighting conditions.
[0031] To achieve white balance correction, the system calculates the proportional difference between the measured image values and the theoretical reference values. The gain coefficients of the three channels are calculated using the following model: ; in, This indicates that all grayscale blocks are in the channel. Theoretical average brightness Channels measured in the image Average brightness on This is the white balance gain coefficient for the corresponding channel. This coefficient reflects the degree of color cast in the current image compared to the standard color model and is used to adjust the overall image color distribution. When the brightness distribution between different grayscale blocks is uneven, the system can introduce regional weights using a weighted average method. To improve the robustness of white balance calculations, the formula is expressed as: ; in, This represents the number of grayscale blocks involved in the calculation. For the first The weight coefficient of each grayscale block.
[0032] The calculated gain coefficient is used for pixel-level color correction of the white-balanced orthophoto. The correction formula is: ; in, Pixels in an orthophoto The original color vector, This is the white balance gain matrix. This is the color value after white balance correction. This calculation is performed point-by-point across the entire pixel range of the image, thus achieving global white balance correction for the entire image.
[0033] Using a physical color calibration board deployed on-site, global white balance correction was performed on the orthophotos, effectively eliminating the overall color cast caused by differences in lighting conditions at different times (such as solar altitude angle and cloud cover) and camera response. The corrected images have high color fidelity and can more realistically reflect the inherent optical characteristics of water bodies, making the RGB color values extracted from the images consistent and comparable, and significantly reducing the interference of ambient light noise on the input of the water quality inversion model.
[0034] S4. Based on the geographic coordinates of the sampling points, extract the RGB color values of the corresponding pixels from the color-fidelity orthophoto mosaic image, and associate and match the RGB color values with the in-situ water quality index data to construct a water quality inversion training dataset. Furthermore, in S4, the specific steps for extracting the RGB color values of corresponding pixels from the color-fidelity orthophoto mosaic image based on the geographic coordinates of the sampling points include: The geographic coordinates of the sampling points are mapped onto the color-fidelity orthophoto mosaic image through coordinate transformation to determine the corresponding center pixel coordinates. A pixel window area of a preset size is defined based on the coordinates of the center pixel. Extract the RGB values of all pixels within the pixel window region; Calculate the average RGB value of all pixels within the pixel window area, and determine the average RGB value as the RGB color value of the corresponding pixel.
[0035] Furthermore, in S4, the specific steps for constructing a water quality inversion training dataset by associating and matching RGB color values with in-situ water quality index data include: Create a data record for each sampling point; The RGB color values corresponding to the sampling points are stored in the data record as input feature vectors; The in-situ water quality index data corresponding to the sampling points are stored in the data record as the expected output label; Data records from all sampling points are collected to generate a water quality inversion training dataset.
[0036] Specifically, the color-preserving true photomosaic image after white balance correction in step S3 is used as input data. Combined with the geographic coordinate information of the field sampling points, the RGB color features corresponding to the spatial location of the sampling points are accurately extracted from the image and matched one by one with the in-situ measured water quality index data to construct the training dataset required for the water quality inversion model.
[0037] First, read the geographic coordinates of the sampling points. The georeferenced information of the imagery (including spatial resolution, projection method, and image origin coordinates) is used to map it to the pixel coordinate system of the orthophoto. The mapping relationship can be expressed as follows: ,in, The geographic coordinates of the upper left corner of the orthophoto are: The images are respectively in Spatial resolution in direction (unit: meters / pixel) For the first The image pixel coordinates corresponding to each sampling point are used. This transformation model achieves precise spatial registration between the sampling points and image pixel locations, ensuring strict consistency between the extracted spectral information and the geographical location.
[0038] After determining the pixel position, the center pixel of the sampling point is used as the reference. Based on the image resolution and the spatial uniformity of the water body, the system defines a pixel window region of a preset size, for example... or The choice of pixel range and window size is determined based on image resolution and water mixing characteristics: a smaller window can be used when the resolution is high and the water is homogeneous to enhance spatial accuracy; if the water spectrum changes gently, a larger window can be used to enhance statistical stability. The set of GB values for all pixels within the window region is extracted and denoted as [missing information]. ,in, This represents the window's side length in pixels.
[0039] To reduce the impact of individual pixel noise and local reflection anomalies, the system calculates the three-channel mean of all pixels within the window area to determine the representative color value for that area. ; in, For the first The average color feature vector of each sampling point reflects the spectral characteristics of the area around the sampling point under the color-fidelity radiograph. This averaging process makes the extracted color features robust to local illumination changes while preserving the main spectral information in the area.
[0040] Subsequently, the extracted color features were fused with in-situ measured water quality indicators. For each sampling point, a complete data record was created, and... Stored in the record as the input feature vector, along with water quality parameters obtained from field sampling (such as chlorophyll). concentration Turbidity Suspended solids concentration electrical conductivity (etc.) are stored as output labels in the record to form corresponding sample pairs: ; Perform the above process on all sampling points to aggregate all sample pairs and generate the training dataset: ; in, For a complete water quality inversion training dataset, For the first The color feature vector of each sampling point This corresponds to the measured water quality parameter vector. This represents the number of sampling points.
[0041] By employing spatial coordinate mapping and local window averaging, representative RGB spectral features corresponding to each sampling point were robustly extracted from color-fidelity images. These features were then correlated one by one with measured water quality data to construct structured "feature-label" sample pairs. The resulting water quality inversion training dataset not only features accurate data correspondence but also exhibits enhanced noise resistance through window averaging. This provides high-quality, highly reliable training samples for machine learning models to learn the complex nonlinear mapping relationship between water color and water quality parameters.
[0042] S5. Use the water quality inversion training dataset to train a machine learning model and establish an inversion model between the RGB optical features of water bodies and water quality indicators. Furthermore, in S5, the specific steps for training a machine learning model using a water quality inversion training dataset to establish an inversion model between the RGB optical features of water bodies and water quality indicators include: The water quality inversion training dataset is divided into a training subset and a validation subset; Select a machine learning model and input the training subset into the machine learning model for training; The model in training is evaluated using a validation subset, and the model hyperparameters are adjusted based on the evaluation results. When the performance metrics of the machine learning model meet the convergence condition, training is stopped, and the model is identified as an inversion model.
[0043] Specifically, step S5 uses the water quality inversion training dataset constructed in step S4 to train a machine learning model to establish a quantitative inversion relationship between the GB optical characteristics of water bodies and water quality indicators, thereby realizing automated prediction from image spectral information to water quality parameters. Through machine learning algorithms, the nonlinear mapping between the color characteristics of different water bodies and measured water quality data is learned, forming a model that can be used for water quality inversion.
[0044] First, the water quality inversion training dataset is prepared. The dataset is divided proportionally into a training subset and a validation subset. Indicates the first RGB color feature vectors extracted from color-fidelity full-photon images at each sampling point This represents the measured water quality index vector at the sampling point, corresponding to chlorophyll. Concentration, turbidity, suspended solids concentration, and conductivity, etc. The training subset is used for model parameter optimization, while the validation subset is used for model performance evaluation and hyperparameter tuning.
[0045] To establish the mapping relationship between GB features and water quality indicators, this invention selects a machine learning model suitable for nonlinear multi-output regression tasks, such as random forest regression. The model approximates the nonlinear functional relationship between RGB and water quality indicators by integrating multiple decision trees. Its mathematical expression is as follows: ; in, This is the water quality index vector predicted by the model. For the number of decision trees, For the first The regression function of the decision tree, This is the set of model parameters (including tree depth, minimum number of sample splits, feature selection strategy, etc.).
[0046] During the training phase, the system uses a training subset of sample pairs. Using the input as input, the model parameters are optimized by minimizing the loss function. The loss function uses mean squared error (MS) to measure the deviation between the predicted results and the measured water quality indicators, and is defined as: ; in, The number of training samples. This refers to the number of water quality indicators (e.g., 4 indicators). For the actual measurement The values of each water quality indicator These are the corresponding indicator values predicted by the model. The model parameters are continuously optimized through iterative processes. , making the loss function The model converges to its minimum value, thus obtaining the optimal model parameters. .
[0047] During training, the system uses a validation subset to evaluate model performance in real time by calculating the coefficient of determination. The model's fitting ability and generalization performance are assessed using metrics such as mean squared error (MSE). The coefficient of determination is defined as: ; in, To verify the sample size, To validate the mean vector of measured water quality indicators in the set. When the model is on the validation subset. When the value tends to stabilize and exceeds the preset threshold (e.g., 0.9), and the verification error decreases to the preset convergence condition, the system stops training and determines the model under the current parameters as the final inversion model.
[0048] The final inversion model Able to adapt to any water body Optical eigenvectors Automatically predict its water quality index estimates: ; in, The water quality predictions output by the model include estimated chlorophyll content. Parameters such as concentration, turbidity, suspended solids concentration, and conductivity.
[0049] Using the constructed training dataset, a quantitative inversion model capable of accurately predicting multiple water quality indicators from the RGB optical features of water bodies was successfully trained. Through a data-driven approach, the model captures the intrinsic correlation between subtle color changes that are difficult for the human eye to perceive and water quality parameters, achieving automated and rapid estimation from simple RGB values to complex water quality parameters, and demonstrating good fitting accuracy and generalization ability.
[0050] S6. Apply the inversion model to the color-fidelity orthophoto mosaic image to generate a spatial distribution map of water quality indicators, and output a warning signal based on the comparison results between the inversion indicator values in the spatial distribution map and the optical warning threshold.
[0051] Furthermore, in S6, the specific steps for applying the inversion model to color-fidelity orthophoto mosaic images to generate spatial distribution maps of water quality indicators include: Traverse the pixels in a color-fidelity orthophoto mosaic image; Extract the RGB color values of the pixels and use them as input features; The input features are fed into the inversion model to obtain the inverted water quality index values corresponding to the pixels. The inverted water quality index values corresponding to all pixels are collected and mapped to the corresponding geographic coordinates to generate a spatial distribution map of the water quality index.
[0052] Furthermore, in S6, the specific steps for outputting the warning signal based on the comparison between the inversion index value in the spatial distribution map and the optical warning threshold include: The inverted water quality index values in the spatial distribution map are compared one by one with the optical early warning threshold; Identify and mark areas where the retrieved water quality index values exceed the optical warning threshold; When a marked area is detected, a warning signal is generated and output.
[0053] Specifically, step S6 uses the water quality inversion model trained in step S5 to predict water quality indicators at the pixel level from the color-preserving true radio mosaic image, generates spatial distribution maps of various water quality parameters, and realizes automatic detection and early warning output of abnormal water conditions based on the comparison between the inversion results and the preset optical early warning threshold. This achieves automatic extraction and intelligent early warning of water quality spatiotemporal distribution information from remote sensing images.
[0054] First, the color-fidelity photomosaic image after geometric and radiometric correction is read, and then every pixel in the image is traversed. For each pixel, the system extracts its three-channel color values. in These represent the red, green, and blue spectral reflectance components of a pixel, respectively. After white balance and color correction, they accurately reflect the optical reflectance characteristics of water. The extracted color vectors... As input features, these are fed into the water quality inversion model established in step S5. In the process, calculate the inversion water quality index value corresponding to this pixel: ; Where is the predicted water quality parameter vector for this pixel, corresponding to the estimated values of chlorophyll concentration, turbidity, suspended solids concentration and conductivity, respectively, and * represents the optimal model parameters that have been trained and converged. The calculation process is performed pixel by pixel across the entire image range to achieve full-domain inversion of water quality indicators.
[0055] After the system completes the water quality index inversion for all pixels, it maps the inversion result of each pixel to its corresponding geographic coordinates based on the image's georeferenced information (including spatial resolution, projection parameters, and origin coordinates). Forming a continuous geospatial dataset: ; in, Given the total number of image pixels, the dataset is transformed into a spatial distribution map of water quality indicators using spatial interpolation and raster visualization techniques, which intuitively displays the variation patterns and abnormal distribution characteristics of each water quality parameter throughout the entire water area.
[0056] After generating the spatial distribution map of water quality, the system enters the optical early warning analysis stage, targeting each type of inverted water quality indicator. The system pre-sets corresponding optical warning thresholds. This threshold is determined based on the optical properties of the water body and ecological safety standards, and is used to identify abnormal or polluted conditions. The system compares the inversion results pixel by pixel with the threshold to determine whether the warning limit has been exceeded. ; in, For the first The pixel in the first Early warning indicator variables under Class I water quality indicators. When When this occurs, it indicates that the inversion results for that region have exceeded the normal range. The system applies this to all... The pixels are clustered and labeled to form a mask layer for the warning area.
[0057] When a marked area is detected, it indicates a potential water quality anomaly or pollution within the monitored water area. An early warning signal is automatically generated and sent to the upper-level monitoring platform. The early warning signal includes information on the geographical extent of the abnormal area, the type of index exceeding the limit, and the magnitude of the exceedance. This information can be visualized on a spatial distribution map using color coding or flashing markers, enabling real-time location and identification of the abnormal area.
[0058] The trained inversion model was applied to orthophotos of the entire wetland, generating thematic maps that visually display the spatial distribution details and variation patterns of various water quality parameters. Finally, by comparing the inversion results with preset thresholds, automatic identification, spatial location, and real-time early warning of areas with abnormal water quality (such as eutrophication and high turbidity areas) were achieved. This elevates traditional discrete point monitoring to continuous, area-based macroscopic monitoring, providing environmental managers with a comprehensive view of the health status of water bodies and precise decision support information.
[0059] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for collecting, investigating, and analyzing wetland environmental conditions, characterized in that, include: S1. Collect UAV imagery data of the wetland study area, and collect in-situ water quality index data and geographic coordinates of multiple sampling points within the wetland; S2. Perform aerial triangulation and orthomorphic mosaicking on the UAV image data to generate an orthomorphic mosaic image. S3. Based on the color calibration plate reference area in the orthophoto mosaic image, perform white balance correction on the orthophoto mosaic image to obtain a color-accurate orthophoto mosaic image. S4. Based on the geographic coordinates of the sampling points, extract the RGB color values of the corresponding pixels from the color-fidelity orthophoto mosaic image, and associate and match the RGB color values with the in-situ water quality index data to construct a water quality inversion training dataset. S5. Use the water quality inversion training dataset to train a machine learning model and establish an inversion model between the RGB optical features of water bodies and water quality indicators. S6. Apply the inversion model to the color-fidelity orthophoto mosaic image to generate a spatial distribution map of water quality indicators, and output a warning signal based on the comparison results of the inversion indicator values in the spatial distribution map and the optical warning threshold.
2. The wetland environmental data collection, investigation, and analysis method according to claim 1, characterized in that, In step S2, performing aerial triangulation and orthomorphic mosaicking on the UAV image data to generate an orthomorphic mosaic image specifically includes the following steps: The drone image data is subjected to quality screening to remove abnormal images; Import the filtered image data, generate a sparse point cloud model through feature matching, and perform bundle adjustment of the regional network. The sparse point cloud model and camera pose parameters are used to perform 3D reconstruction, generating a dense point cloud and texture model. The images are geometrically corrected and color balanced, and overlapping areas are joined together to create an orthophoto mosaic.
3. The wetland environmental data collection, investigation, and analysis method according to claim 1, characterized in that, In step S3, based on the color calibration plate reference area in the orthophoto mosaic image, white balance correction is performed on the orthophoto mosaic image to obtain a color-accurate orthophoto mosaic image. This specifically includes the following steps: Identify and locate the color calibration plate reference area in the orthophoto mosaic image; The average RGB color value is extracted from the grayscale block of the reference area of the color calibration plate as the image measurement value; The RGB color values of the grayscale block are obtained as theoretical reference values; Based on the difference between the image measurement value and the theoretical reference value, the white balance gain coefficients of the R, G, and B channels are calculated. The white balance gain coefficient is applied to the orthophoto mosaic image to obtain a color-fidelity orthophoto mosaic image.
4. The wetland environmental data collection, investigation, and analysis method according to claim 1, characterized in that, In step S4, the specific steps for extracting the RGB color values of corresponding pixels from the color-fidelity orthophoto mosaic image based on the geographic coordinates of the sampling points include: The geographic coordinates of the sampling points are mapped onto the color-fidelity orthophoto mosaic image through coordinate transformation to determine the corresponding center pixel coordinates. Using the coordinates of the center pixel as a reference, a pixel window region of a preset size is defined; Extract the RGB values of all pixels within the pixel window region; Calculate the average RGB value of all pixels within the pixel window area, and determine the average RGB value as the RGB color value of the corresponding pixel.
5. The wetland environmental data collection, investigation, and analysis method according to claim 1, characterized in that, In step S4, the specific steps for associating and matching the RGB color values with the in-situ water quality index data to construct a water quality inversion training dataset include: For each of the sampling points, create a data record; The RGB color value corresponding to the sampling point is stored in the data record as an input feature vector; The in-situ water quality index data corresponding to the sampling point are stored in the data record as the expected output label; Data records from all the sampling points are collected to generate a water quality inversion training dataset.
6. The wetland environmental data collection, investigation, and analysis method according to claim 1, characterized in that, In step S5, the specific steps for training a machine learning model using the water quality inversion training dataset and establishing an inversion model between the RGB optical features of water bodies and water quality indicators include: The water quality inversion training dataset is divided into a training subset and a validation subset; A machine learning model is selected, and the training subset is input into the machine learning model for training; The model in training is evaluated using the validation subset, and the model hyperparameters are adjusted based on the evaluation results. When the performance metrics of the machine learning model meet the convergence condition, training is stopped, and the model is identified as an inversion model.
7. The wetland environmental data collection, investigation, and analysis method according to claim 1, characterized in that, In step S6, the specific steps of applying the inversion model to the color-fidelity orthophoto mosaic image to generate a spatial distribution map of water quality indicators include: Traverse the pixels in the color-fidelity orthophoto mosaic image; Extract the RGB color values of the pixels and use them as input features; The input features are input into the inversion model to obtain the inverted water quality index value corresponding to the pixel; The retrieved water quality index values corresponding to all pixels are collected and mapped to the corresponding geographic coordinates to generate a spatial distribution map of the water quality index.
8. The method for collecting, investigating, and analyzing wetland environmental conditions according to claim 1, characterized in that, In step S6, the specific steps for outputting a warning signal based on the comparison between the inversion index value in the spatial distribution map and the optical warning threshold include: The inverted water quality index values in the spatial distribution map are compared one by one with the optical early warning threshold; Identify and mark areas where the retrieved water quality index values exceed the optical warning threshold; When the marked area is detected, a warning signal is generated and output.