Method for measuring sand content of water body based on unmanned aerial vehicle

Abstract

The invention discloses a method for measuring the sand content of a water body based on an unmanned aerial vehicle. The method comprises the following steps: 1, establishing an inference model from a digital quantized value of a target water body image to the sand content of the target water body; secondly, the unmanned aerial vehicle flies along the target water body, collects images of different areas of the target water body and transmits the images to the remote computer; thirdly, manually calibrating a water body area in each image acquired by the unmanned aerial vehicle; fourthly, the computer calculates the average value DN1 of the digital quantized values of all the pixels according to the calibrated water body areas in all the images; and fifthly, through the reasoning model established in the first step, the computer obtains the sand content corresponding to each image acquired by the unmanned aerial vehicle according to the DN1 value corresponding to each image acquired by the unmanned aerial vehicle. Through model driving, unmanned aerial vehicle dynamic acquisition and data processing, efficient, accurate and low-cost monitoring of the sand content of the water body is realized, and the system is suitable for river management requirements in a large-range and complex environment, and provides reliable technical support for water resource protection, disaster early warning and ecological restoration.

Description

Technical Field

[0001] The present invention relates to the field of environmental monitoring technology in the advanced environmental protection industry, in particular to the technology of monitoring the sediment content of water bodies by unmanned aerial vehicles (UAVs), which can be used to manufacture special UAVs for environmental protection. Background Art

[0002] 1. The significance of sediment content measurement

[0003] Suspended sediment has a profound impact on the physical, chemical and biological processes of water bodies, and real-time monitoring of the sediment content in river water bodies is crucial.

[0004] The Yellow River, China's second-longest river, is also one of the most sediment-laden rivers in the world. It flows through several Chinese provinces before emptying into the Bohai Sea. The Yellow River's high sediment content is primarily due to the geological and climatic conditions of the region it flows through, particularly soil erosion on the Loess Plateau. Large amounts of sediment are washed into the Yellow River annually, contributing to its extremely high sediment content.

[0005] High sediment concentrations have significant impacts on the Yellow River's ecosystem and downstream activities such as agriculture and shipping. For example, sediment accumulation can cause riverbed uplift, increasing flood risk and placing pressure on water conservancy projects and irrigation systems. Furthermore, sediment deposition affects the capacity and lifespan of reservoirs.

[0006] Measuring the sediment content of the Yellow River has important scientific and ecological significance, mainly reflected in the following aspects:

[0007] 1. Ensure the safety and efficiency of water conservancy projects

[0008] The Yellow River is famous for its high sediment content, and siltation is a core issue threatening the safety of water conservancy projects.

[0009] In terms of reservoir sedimentation management, reservoirs such as the Sanmenxia and Xiaolangdi reservoirs have long suffered from sediment accumulation, resulting in reduced reservoir capacity and directly impacting flood control, power generation, and water supply. Sediment concentration monitoring can optimize reservoir operations (such as water and sediment regulation), reduce sediment deposition, and extend the life of the project.

[0010] Regarding river stability, sediment accumulation can raise the riverbed, creating a "hanging river" (e.g., in the lower Yellow River), increasing the risk of levee breaches. Real-time monitoring of sediment concentrations can provide data support for river dredging and levee reinforcement.

[0011] 2. Ecological protection and water quality management

[0012] Sediment is not only a physical carrier, but is also closely related to the chemical and biological processes of water bodies.

[0013] In terms of water turbidity control, high sediment concentrations cause water turbidity, affecting photosynthesis of aquatic organisms and destroying fish spawning grounds and benthic habitats. Monitoring sediment concentrations can guide ecological restoration projects (such as wetland protection and vegetation restoration).

[0014] In the study of pollutant migration, sediment absorbs pollutants such as heavy metals and pesticides. The sediment content data can be used to assess the risk of pollutant diffusion and provide a basis for pollution prevention and control.

[0015] Among agricultural non-point source pollution in the Yellow River Basin, nitrogen and phosphorus carried by sediment are one of the main sources of eutrophication of water bodies. Accurate monitoring of sediment content helps in calculating pollution load.

[0016] 3. Supporting efficient use of water resources

[0017] The Yellow River is an important source of water supply in northern China, and its sediment content directly affects the efficiency of water resource utilization.

[0018] When it comes to irrigation system optimization, high sediment concentrations can easily clog irrigation canals, increasing maintenance costs. By monitoring sediment concentrations, water diversion timing can be optimized (e.g., avoiding periods of high sediment concentrations) to reduce the impact of sediment on agricultural irrigation.

[0019] In terms of industrial water management, thermal power, metallurgy and other industries have high requirements for water quality. Sediment content data can guide the selection of water treatment processes and reduce equipment wear.

[0020] The Yellow River irrigation areas in Ningxia, Inner Mongolia and other places have reduced the cost of channel dredging by tens of millions of yuan each year by adjusting water diversion plans through sediment content monitoring.

[0021] 4. Disaster early warning, prevention and mitigation

[0022] Sudden changes in sand content are potential signals of geological disasters.

[0023] In terms of landslide warning, a surge in sand content during heavy rain may reflect increased mountain erosion, which can be used to predict landslide risks in combination with topographic data.

[0024] In terms of flood risk prediction, floods with high sediment content are more destructive (for example, floods carrying sediment have greater impact), and real-time monitoring can provide support for flood control decisions.

[0025] 5. Scientific research and policy making

[0026] Long-term sediment concentration data are the basis for studying the environmental evolution of the Yellow River Basin.

[0027] In climate change research, changes in sediment content can reflect the long-term impact of climate factors such as precipitation and vegetation cover.

[0028] In terms of soil and water conservation assessment, the effects of projects such as returning farmland to forests and terrace construction can be quantitatively evaluated through changes in sand content.

[0029] In terms of international river management reference, the Yellow River's sediment management experience is of reference significance to the world's sandy rivers (such as the Ganges and the Mississippi River).

[0030] Measuring the Yellow River's sediment concentration is not only a technical requirement but also a strategic task to ensure basin safety and ecological health. Precise monitoring supports multiple goals, including controlling sediment, preventing disasters, ensuring resource availability, and protecting the ecosystem, providing scientific support for the long-term development of the Yellow River basin and the nation as a whole.

[0031] 2. Introduction to existing water sediment content measurement technology

[0032] 1. Traditional measurement methods

[0033] Direct measurement method (such as filtration method, drying method, specific gravity method), the steps are as follows:

[0034] On-site sampling: collecting sand-laden water samples from water bodies (e.g. using sampling bottles or containers).

[0035] Dewatering treatment;

[0036] Filtration method: Separate the sediment from the water through filter paper and retain the sediment.

[0037] Drying method: Heat the water sample until the water evaporates and weigh the remaining sediment.

[0038] Specific gravity method: Place the water sample into a specific gravity bottle, weigh it, and then calculate the mass of the sediment using the formula.

[0039] Calculation of sediment content: Calculate sediment content (kg / m 3 ).

[0040] Disadvantages: cumbersome operation, environmental damage, low efficiency, and difficulty in rapid measurement of multiple monitoring points.

[0041] 2. Indirect measurement method

[0042] Vibration method: Flow sediment-laden water through a vibrating tube and measure the change in resonant frequency. Sediment concentration is inverted using a frequency-sediment relationship model.

[0043] Disadvantages: Affected by temperature and fluid viscosity; low efficiency, and difficult to quickly measure multiple monitoring points.

[0044] Isotope method:

[0045] This method uses gamma rays or neutrons to penetrate water and measure the attenuation of the radiation. The sediment content is calculated based on the relationship between the attenuation characteristics and the sediment concentration. Disadvantage: There is a risk of radioactive safety.

[0046] Capacitance method:

[0047] Measures the change in the dielectric constant of sediment-laden water. Inverts sediment concentration using a "concentration-capacitance" model. Disadvantage: Decreased linearity at high concentrations.

[0048] Optical method:

[0049] The laser or spectral signal is transmitted through the water and the light intensity attenuation or backscattered signal is measured.

[0050] The sediment content is calculated based on the statistical model of optical signals and sediment content.

[0051] Disadvantages: Affected by water turbidity and bubbles.

[0052] Acoustic method:

[0053] Ultrasonic waves are emitted and the attenuation or frequency shift of the sound waves in suspended particles is measured.

[0054] The sediment content is inverted through the relationship model between acoustic characteristics and sediment content.

[0055] Disadvantages: Need to compensate for flow rate and temperature changes.

[0056] Chinese patent CN201911314905.5 discloses a method for measuring the suspended sediment transport rate based on the fluorescence principle. In the later experimental process, it is only necessary to take a photo of the water body and then extract the grayscale of the photo (corresponding to the digital quantization value in the present invention). The sediment content of the water body can be quickly obtained through the relationship between the sediment content and the grayscale of the photo. Although it is relatively close to the present invention, it relies on fluorescence technology and requires the use of sediment coated with fluorescent paint to obtain the sediment content distribution on the cross section through the fluorescence principle. Since sediment coated with fluorescent paint is required, it is only suitable for experiments and not for natural water bodies.

[0057] The reason why it must use fluorescence technology is that it is impossible to find the pattern between the digital quantization value (grayscale value) of the water image and the sand content of the water body without using fluorescence technology, so it can only use fluorescence technology.

[0058] Furthermore, remote sensing technology used to monitor sediment concentrations in water bodies faces multiple interferences, including atmospheric conditions and the river basin environment, and lacks research on dynamic water bodies. Therefore, there is an urgent need to explore more cost-effective, efficient, and adaptable monitoring technologies. Summary of the Invention

[0059] The purpose of the present invention is to provide a method for measuring the sediment content in water bodies based on drones, so as to complete the task of monitoring the sediment content in water bodies at a lower cost and higher efficiency.

[0060] To achieve the above objectives, the method for measuring sediment concentration in water bodies based on drones of the present invention is carried out in the following steps:

[0061] The first is to establish an inference model from the digital quantification value of the target water body image to the sediment content of the target water body;

[0062] The second is that the drone flies along the target water body and collects images of different areas of the target water body and transmits the images back to a remote computer;

[0063] The third is to manually demarcate the water area in each image collected by the drone;

[0064] Fourth, the computer calculates the average value DN1 of the digital quantization value of each pixel for the marked water area in each image;

[0065] The fifth is that through the inference model established in the first step, the computer obtains the sand content corresponding to each image collected by the drone based on the DN1 value corresponding to each image collected by the drone.

[0066] The first step includes the following sub-steps:

[0067] First, for the target water body, confirm its daily sediment content range [min1, max1] and the sediment content range during water and sediment regulation (max1, max2] by measuring or querying historical data, where min1, max1, and max2 are all in kilograms per cubic meter;

[0068] The second is to establish a radiometric calibration function to determine the reflectivity of water images collected by UAVs;

[0069] The third is to establish a benchmark function to establish the functional relationship between reflectivity and sediment content of the target water body.

[0070] In the first sub-step of the first step, min1=0, max1=20, and max2=300.

[0071] The second sub-step of the first step is specifically:

[0072] (1) Hardware preparation: A drone equipped with a multispectral camera was used. A light intensity sensor was integrated on the top of the drone to record the spectral irradiance of each band during shooting. Three calibration plates (15 cm × 15 cm in size) were used: black, gray, and white. The standard reflectance of each band was pre-measured by a hyperspectral instrument.

[0073] (2) Environmental requirements: Good outdoor visibility to avoid shadows or debris covering the calibration plate; the drone's flight altitude is fixed at 2 meters to ensure that the entire calibration plate is fully captured by the multispectral camera;

[0074] (3) Data collection process:

[0075] (1) Calibration board layout: Place three calibration boards (black, gray, and white) horizontally on an open ground. Keep the surface clean and free of debris to avoid reflection interference. Use a level to ensure that the calibration boards are parallel to the ground to reduce geometric distortion.

[0076] (2) Set the camera parameters; (3) Collect drone data;

[0077] Control the drone to hover 2 meters above the calibration plate and ensure that the calibration plate is in the center of the image;

[0078] Capture multispectral images, generating six images per capture: one RGB visible light image, four single-band images, and one four-band composite image; the four single-band images include a green light image, a red light image, a red-edge image, and a near-infrared image;

[0079] Time series acquisition: repeated shooting at different times of the day to cover different sun altitudes and light intensities;

[0080] (IV) Image preprocessing:

[0081] (1) Data classification and storage: Classify and store images by different bands, extract metadata to record shooting time, light intensity sensor data, and camera parameters;

[0082] (2) DN value correction: The digital quantization value is the DN value. According to the black level value and gain value in each image metadata, the DN value is subjected to black level correction and gain compensation, thereby eliminating the sensor dark current noise and electronic amplification noise;

[0083] (3) Distortion correction: Based on the camera calibration parameters, use MATLAB or Python OpenCV tools to perform geometric correction to eliminate image edge deformation;

[0084] (5) Establishing radiation calibration function:

[0085] (1) Extract the DN value of the calibration plate area;

[0086] In each single-band image, the average DN value was calculated;

[0087] (2) Fitting the reflectance-DN relationship model;

[0088] Data matching, associating the known reflectivity of the calibration plate with the corresponding DN value;

[0089] The linear function, exponential function and power function are used to fit the DN value and reflectivity. The mathematical expression is:

[0090] The linear function is: R(γ)=G γ ×DN”(γ)+B γ ;

[0091] The exponential function is:

[0092] The power function is:

[0093] In the formula, γ represents one of the five bands, R(γ) is the reflectivity of the corresponding radiation calibration plate in the γ band image, and G γ and B γ is the parameter value, DN”(γ) is the average DN value of the corresponding radiation calibration plate in the corrected γ band image;

[0094] To establish a quantitative relationship between the DN value and reflectivity of images captured by a spectral camera, the average DN value and reflectivity data of the radiometric calibration plate were imported into MATLAB's Curve Fitting Tool for linear, exponential, and power function fitting analysis. To evaluate the accuracy and reliability of different fitting models, four key statistical parameters were used to analyze the fitting results: residual sum of squares, coefficient of determination, adjusted coefficient of determination, and root mean square error.

[0095] Through fitting and analysis, the final radiation calibration function is determined as follows:

[0096] The green light image is matched with the first radiometric calibration function;

[0097] The red edge image is matched with the second radiometric calibration function;

[0098] The first radiation calibration function is: RG = 0.1235*exp(3.043e-05*x); where RG is the normalized reflectance of the green light band, and x in this formula is the average DN value of the specific green light image after DN value correction;

[0099] The second radiation calibration function is: RRE = 5.67e-07*x^1.289; where RRE is the normalized reflectance of the red light band, and x in this formula is the average DN value of the specific red edge image after DN value correction.

[0100] The third sub-step of the first step is specifically:

[0101] [min1, max1] is called the first sediment concentration interval, and (max1, max2] is called the second sediment concentration interval. After fitting and evaluation, the final benchmark functions include the green light benchmark function and the red edge benchmark function.

[0102] The green light reference function is: for the first sediment concentration interval, f(x) = 1.528e+17*x^19.38; for the second sediment concentration interval, f(x) = 1.031e+04*x-1551;

[0103] The red edge benchmark function is: for the first sediment concentration interval, f(x) = 2599*x^1.595-1.697; for the second sediment concentration interval, f(x) = 3400*x^1.202-78.31;

[0104] Wherein, f(x) is the sediment content of the water body; x in the first benchmark function is the standardized reflectance of the specific green light image calculated by the first radiation calibration function, and x in the first benchmark function is the standardized reflectance of the specific red edge image calculated by the second radiation calibration function.

[0105] min1=0,max1=20,max2=300,the unit is kilograms per cubic meter; 3 ,25kg / m 3 ] is defined as the sediment content transition interval; in the sediment content transition interval, when calculating the sediment content for a specific water body photo, the weighted average of the green light benchmark function and the red edge benchmark function is used as the final water body sediment content; when it is lower than the sediment content transition interval, only the green light benchmark function is used to calculate the water body sediment content; when it is higher than the sediment content transition interval, only the red edge benchmark function is used to calculate the water body sediment content.

[0106] The present invention has the following advantages:

[0107] This paper proposes a technical approach to establish an inference model from the digital quantization value of the target water body image to the target water body sediment content, and to use drones to collect natural water body images and then identify the natural water body sediment content. Overall, this approach has the following advantages:

[0108] 1. High efficiency and wide coverage of natural water bodies.

[0109] Drones can fly quickly along the target water body and obtain high-resolution images of a large area of water in a short period of time, significantly improving monitoring efficiency and avoiding the defects of traditional single-point sampling that is time-consuming and has limited coverage.

[0110] It can adapt to complex terrain (such as mountainous areas and rapid river sections) without relying on fixed monitoring sites, and is particularly suitable for areas that are difficult to reach with traditional equipment.

[0111] 2. High precision and non-contact measurement.

[0112] By combining digital quantification (DN) values with a quantitative inference model of sediment concentration, human error is reduced, resulting in higher accuracy. Drones flying at low altitudes (e.g., 2 meters) are virtually unaffected by atmospheric interference, and their spectral signals are closer to the actual reflectance characteristics of water bodies than satellite remote sensing.

[0113] 3. High data consistency.

[0114] The water areas in the image were manually calibrated, removing interfering objects like waves and floating objects. Only the pure water areas were retained for calculation of the average grayscale, eliminating interference from non-target objects and ensuring that the digital quantization value only reflects the sediment content in the water. The average DN value of all pixels within the calibrated water area in the image was used as input to reduce the impact of local noise on the results and enhance data reliability.

[0115] 4. Automation and real-time performance.

[0116] Aside from demarcating water areas and monitoring drones, the system requires virtually no human intervention, reducing operational complexity. This leverages the low cost and speed of manual water area demarcation while minimizing human error. Real-time image data transmitted via wireless communication modules, combined with rapid computer processing, can generate sediment concentration maps in a fraction of a second, supporting emergency decision-making, such as flood warnings.

[0117] 5. Significant cost-effectiveness

[0118] Monitoring can be accomplished using consumer-grade drones equipped with multispectral cameras, eliminating the need for expensive specialized equipment. Drones are flexible and eliminate the need for extensive on-site sampling or installation of measurement equipment at the water body, reducing long-term operational costs.

[0119] 6. Adapt to complex scenarios

[0120] It is suitable for complex water environments such as high turbidity and turbulence, and maintains stable performance in scenarios where traditional optical methods are susceptible to interference. By optimizing the inference model (incorporating multi-band data), it can be expanded to simultaneously monitor other water quality parameters (such as suspended matter concentration and chlorophyll a).

[0121] In summary, this invention achieves efficient, accurate, and low-cost monitoring of water sediment content through "model-driven + UAV dynamic collection + intelligent data processing". It is particularly suitable for river management needs in large-scale and complex environments, and provides reliable technical support for water resource protection, disaster warning, and ecological restoration.

[0122] The sub-steps of the first step reflect the inventor's innovative understanding of the regularity of the relationship between the digital quantization value of the water body image and the sediment content of the water body without using fluorescence technology.

[0123] 1. It's impossible to establish a direct correlation between the digital quantization value of a water body image and its sediment concentration. This is because the average digital quantization value (DN) is affected by factors such as lighting conditions, shooting angle, and environmental reflections (such as water surface reflections and floating objects). Directly mapping the DN to sediment concentration can lead to amplified errors. For example, if a water body with the same sediment concentration is photographed on a cloudy day and a sunny day, the DN value will differ significantly, causing the inversion result to deviate from the true value.

[0124] This is also the important reason why CN201911314905.5 can only use sediment coated with fluorescent paint to establish a direct correlation between the grayscale of the water image and the sand content of the water.

[0125] 2. The relationship between sediment content and digital quantification value is nonlinear in different concentration ranges (for example, reflectivity changes sensitively at low concentrations and tends to saturation at high concentrations). A single model cannot accurately cover the entire range.

[0126] Based on the above-mentioned innovative regularity, the inventors constructed three sub-steps of the first step, reflecting the following technical considerations of the inventors:

[0127] 1. Radiometric calibration eliminates environmental interference. Calibration using black, gray, and white calibration plates establishes a quantitative relationship between DN and reflectivity, eliminating the effects of varying illumination. Raw digital values are converted to standardized reflectivity, ensuring data comparability across time and weather conditions.

[0128] By converting digital quantization values into reflectance through a radiation calibration function, external interference factors (such as changes in light intensity and differences in sensor response) can be removed, making the data closer to the essential optical properties of the water body.

[0129] 2. Multispectral band optimization and segmented modeling.

[0130] The green light (G) and red edge (RE) bands are preferred, as their reflectivity has been compared with other bands by the applicant and is found to be more sensitive to changes in sediment content.

[0131] The model is established in different intervals: the power function of the green light band is used for low sediment content, and the power function of the red edge band after translation is used for high sediment content. The complementary characteristics of the green light (G) and red edge (RE) bands discovered by the inventors in different sediment content intervals are utilized.

[0132] 3. Targeted modeling: The optimal function form is used for different bands and concentration ranges to fit the actual physical relationship and avoid the problem of no overall regularity in all bands and concentrations.

[0133] 4. The physical correlation between sediment content and reflectivity is more direct than digital quantification. Sediment particles alter water reflectivity through scattering and absorption, and the relationship between reflectivity and sediment content is closer to physical laws (e.g., based on Mie scattering theory). Therefore, using reflectivity as an intermediate parameter between the average digital quantification value of water images and sediment content allows for the construction of a more accurate quantitative model.

[0134] 5. Multi-spectral collaborative analysis.

[0135] Reflectance responds differently in different wavelengths of light. Creating composite features by combining multispectral reflectance can suppress noise interference from a single wavelength. Using spectral images from different wavelengths for water bodies with varying sediment concentrations improves model robustness and accuracy.

[0136] By using the specific techniques in the first step, the full range can be covered, the nonlinearity of a single model can be resolved, and the inversion accuracy in high and low concentration zones can be improved.

[0137] Direct mapping of sediment concentrations based solely on average digital quantization suffers from inherent drawbacks such as high environmental sensitivity, neglect of nonlinearity, and poor interference immunity. The comprehensive approach presented in this paper significantly improves accuracy and reliability through radiometric calibration, multispectral optimization, segmented modeling, and data cleaning, offering a superior and more reliable technical approach. The weighted average method avoids model switching errors and provides a smooth transition in calculated results. BRIEF DESCRIPTION OF THE DRAWINGS

[0138] Figure 1 This is an image of the Yellow River surface taken by a drone; it is easy for staff to Figure 1 The water area is identified from the areas of foam and floating objects, and the water area is calibrated to eliminate the influence of foam and floating objects. When calibrating the water area, it is not necessary to distinguish the water area from the foreign object area at the pixel level; it is only necessary to calibrate the majority of the water area in the image.

[0139] Figure 2 It is a sand chart taken from the experiment sand.

[0140] Figure 3 This is an overview of the outdoor experiment.

[0141] Figure 4 Example of capturing images for a multispectral camera.

[0142] Figure 5 It is a scatter plot of reflectivity and sediment content corresponding to each other in two sediment content intervals.

[0143] Figure 6 It is the green light benchmark function graph.

[0144] Figure 7 It is the red edge benchmark function graph.

[0145] Figure 8 It is a comparison chart of the original image, the image after DN value correction and the image after distortion correction.

[0146] Figure 9 This is a schematic diagram of the green (G) band radiation calibration fitting function. DETAILED DESCRIPTION

[0147] like Figures 1 to 9As shown, the method for measuring sediment content in water bodies based on drones of the present invention is carried out in the following steps:

[0148] The first is to establish an inference model from the digital quantization value of the target water body image to the sediment content of the target water body; the inference model includes a radiation calibration function and a benchmark function, wherein the radiation calibration function establishes a quantitative relationship between the DN value and the reflectivity through calibration of the calibration plate, and the benchmark function is based on multispectral band segmentation modeling, which includes the complementary characteristics of the green light and red edge bands.

[0149] The second is that the UAV flies along the target water body and collects images of different areas of the target water body, and transmits the images back to the remote computer (the wireless communication module can be used to realize simultaneous image collection and transmission, or the UAV storage module can be directly read after the UAV returns to obtain images of different areas of the target water body);

[0150] The third step is to manually demarcate the water area in each image collected by the drone (removing objects such as waves and floating objects in the image to prevent these objects with significantly different grayscale from the water from affecting the authenticity of the average digital quantization value of the water body);

[0151] Fourth, the computer calculates the average value of the digital quantization value of each pixel (i.e. the average value of the DN value of each pixel) DN1 for the marked water area in each image;

[0152] The fifth is that through the inference model established in the first step, the computer obtains the sand content corresponding to each image collected by the drone based on the DN1 value corresponding to each image collected by the drone.

[0153] This paper proposes a technical approach to establish an inference model from the digital quantization value of the target water body image to the target water body sediment content, and to use drones to collect natural water body images and then identify the natural water body sediment content. Overall, this approach has the following advantages:

[0154] 1. High efficiency and wide coverage of natural water bodies.

[0155] Drones can fly quickly along the target water body and obtain high-resolution images of a large area of water in a short period of time, significantly improving monitoring efficiency and avoiding the defects of traditional single-point sampling that is time-consuming and has limited coverage.

[0156] It can adapt to complex terrain (such as mountainous areas and rapid river sections) without relying on fixed monitoring sites, and is particularly suitable for areas that are difficult to reach with traditional equipment.

[0157] 2. High precision and non-contact measurement.

[0158] By combining digital quantification (DN) values with a quantitative inference model of sediment concentration, human error is reduced, resulting in higher accuracy. Drones flying at low altitudes (e.g., 2 meters) are virtually unaffected by atmospheric interference, and their spectral signals are closer to the actual reflectance characteristics of water bodies than satellite remote sensing.

[0159] 3. High data consistency.

[0160] The water areas in the image were manually calibrated, removing interfering objects like waves and floating objects. Only the pure water areas were retained for calculation of the average grayscale, eliminating interference from non-target objects and ensuring that the digital quantization value only reflects the sediment content in the water. The average DN value of all pixels within the calibrated water area in the image was used as input to reduce the impact of local noise on the results and enhance data reliability.

[0161] 4. Automation and real-time performance.

[0162] Aside from demarcating water areas and monitoring drones, the system requires virtually no human intervention, reducing operational complexity. This leverages the low cost and speed of manual water area demarcation while minimizing human error. Real-time image data transmitted via wireless communication modules, combined with rapid computer processing, can generate sediment concentration maps in a fraction of a second, supporting emergency decision-making, such as flood warnings.

[0163] 5. Significant cost-effectiveness

[0164] Monitoring can be accomplished using consumer-grade drones equipped with multispectral cameras, eliminating the need for expensive specialized equipment. Drones are flexible and eliminate the need for extensive on-site sampling or installation of measurement equipment at the water body, reducing long-term operational costs.

[0165] 6. Adapt to complex scenarios

[0166] It is suitable for complex water environments such as high turbidity and turbulence, and maintains stable performance in scenarios where traditional optical methods are susceptible to interference. By optimizing the inference model (incorporating multi-band data), it can be expanded to simultaneously monitor other water quality parameters (such as suspended matter concentration and chlorophyll a).

[0167] Table 1: Performance comparison between the present invention and the prior art

[0168] index Traditional methods (such as sampling + laboratory analysis) UAV multispectral methods Coverage Single point or local area Large-scale, full-basin coverage Timeliness Lag (sampling, transportation, analysis required) Near real-time Human dependence High (on-site sampling personnel required) Low (mainly automated processing) Environmental adaptability Limited by the accessibility of sampling points Adapt to complex terrain and dynamic waters Long-term monitoring costs High (repeated sampling and laboratory costs) Low (one-time investment, long-term reuse)

[0169] In summary, this invention achieves efficient, accurate, and low-cost monitoring of water sediment content through "model-driven + UAV dynamic collection + intelligent data processing". It is particularly suitable for river management needs in large-scale and complex environments, and provides reliable technical support for water resource protection, disaster warning, and ecological restoration.

[0170] The first step includes the following sub-steps:

[0171] The first is to determine the daily sediment content range [min1, max1] and the sediment content range during water and sediment regulation (max1, max2] for target water bodies (such as the Yellow River, Yangtze River, etc.) by measuring or querying historical data. The units of min1, max1, and max2 are all kilograms per cubic meter (kg / m 3 );

[0172] The second is to establish a radiometric calibration function to determine the reflectivity of water images collected by UAVs;

[0173] The third is to establish a benchmark function to establish the functional relationship between reflectivity and sediment content of the target water body.

[0174] The sub-steps of the first step reflect the inventor's innovative understanding of the regularity of the relationship between the digital quantization value of the water body image and the sediment content of the water body without using fluorescence technology.

[0175] 1. It's impossible to establish a direct correlation between the digital quantization value of a water body image and its sediment concentration. This is because the average digital quantization value (DN) is affected by factors such as lighting conditions, shooting angle, and environmental reflections (such as water surface reflections and floating objects). Directly mapping the DN to sediment concentration can lead to amplified errors. For example, if a water body with the same sediment concentration is photographed on a cloudy day and a sunny day, the DN value will differ significantly, causing the inversion result to deviate from the true value.

[0176] This is also the important reason why CN201911314905.5 can only use sediment coated with fluorescent paint to establish a direct correlation between the grayscale of the water image and the sand content of the water.

[0177] 2. The relationship between sediment content and digital quantification value is nonlinear in different concentration ranges (for example, reflectivity changes sensitively at low concentrations and tends to saturation at high concentrations). A single model cannot accurately cover the entire range.

[0178] Based on the above-mentioned innovative regularity, the inventors constructed three sub-steps of the first step, reflecting the following technical considerations of the inventors:

[0179] 1. Radiometric calibration eliminates environmental interference. Calibration using black, gray, and white calibration plates establishes a quantitative relationship between DN and reflectivity, eliminating the effects of varying illumination. Raw digital values are converted to standardized reflectivity, ensuring data comparability across time and weather conditions.

[0180] By converting digital quantization values into reflectance through a radiation calibration function, external interference factors (such as changes in light intensity and differences in sensor response) can be removed, making the data closer to the essential optical properties of the water body.

[0181] 2. Multispectral band optimization and segmented modeling.

[0182] The green light (G) and red edge (RE) bands are preferred, as their reflectivity has been compared with other bands by the applicant and is found to be more sensitive to changes in sediment content.

[0183] The model is established in different intervals: the power function of the green light band is used for low sediment content, and the power function of the red edge band after translation is used for high sediment content. The complementary characteristics of the green light (G) and red edge (RE) bands discovered by the inventors in different sediment content intervals are utilized.

[0184] 3. Targeted modeling: The optimal function form is used for different bands and concentration ranges to fit the actual physical relationship and avoid the problem of no overall regularity in all bands and concentrations.

[0185] 4. The physical correlation between sediment content and reflectivity is more direct than digital quantification. Sediment particles alter water reflectivity through scattering and absorption, and the relationship between reflectivity and sediment content is closer to physical laws (e.g., based on Mie scattering theory). Therefore, using reflectivity as an intermediate parameter between the average digital quantification value of water images and sediment content allows for the construction of a more accurate quantitative model.

[0186] 5. Multi-spectral collaborative analysis.

[0187] Reflectance responds differently in different wavelengths of light. Creating composite features by combining multispectral reflectance can suppress noise interference from a single wavelength. Using spectral images from different wavelengths for water bodies with varying sediment concentrations improves model robustness and accuracy.

[0188] By using the specific techniques in the first step, the full range can be covered, the nonlinearity of a single model can be resolved, and the inversion accuracy in high and low concentration zones can be improved.

[0189] Table 2 Comparison of sand content calculated by the present invention and the simple average digital quantization value method

[0190]

[0191] Direct mapping of sediment content that relies solely on average digital quantization values has inherent defects such as strong environmental sensitivity, neglect of nonlinearity, and poor anti-interference. The comprehensive method of the present invention significantly improves accuracy and reliability through radiation calibration, multi-spectral optimization, segmented modeling and data cleaning, and is a better and more reliable technical path.

[0192] In the first sub-step of the first step, for the Yellow River, min1=0, max1=20, and max2=300.

[0193] The determination of the various parameters of the sediment content range is based on the natural distribution characteristics of the sediment content in the Yellow River Basin. According to long-term monitoring data from the Huayuankou Hydrological Station on the Yellow River, daily sediment content is usually stable in the range of [0, 20], but during water and sediment regulation, the sediment content can surge to the range of [20, 300].3 It is the critical value for distinguishing daily low sediment concentration from high sediment concentration during the water and sediment regulation period, and reflects the typical characteristics of the dynamic changes in the Yellow River sediment.

[0194] The second sub-step of the first step is carried out through a radiation calibration experiment, specifically:

[0195] (I) Hardware Preparation: Use an unmanned aerial vehicle (UAV) equipped with a (16-bit) multispectral camera, preferably a DJI Mavic 3 UAV. The multispectral camera carries four wavelength bands: green (G, 560 nm), red (R, 650 nm), red edge (RE, 730 nm), and near-infrared (NIR, 840 nm). A light intensity sensor is integrated on the top of the UAV to record the spectral irradiance of each wavelength band during shooting. Three calibration plates (15 cm × 15 cm in size, black, gray, and white) are provided. The standard reflectance of each wavelength band is pre-measured using a hyperspectral instrument (e.g., PSR-1100) (obtained in Table 4).

[0196] (2) Environmental requirements: Outdoor visibility conditions must be good (cloudless and stable lighting) to avoid shadows blocking the calibration plate or debris covering the calibration plate; the drone's flight altitude must be fixed at 2 meters to ensure that the entire calibration plate is fully captured by the multispectral camera;

[0197] (3) Data collection process:

[0198] (1) Calibration board layout: Place three calibration boards (black, gray, and white) horizontally on an open ground. Keep the surface clean and free of debris to avoid reflection interference. Use a level to ensure that the calibration boards are parallel to the ground to reduce geometric distortion.

[0199] (2) Set the camera parameters. The specific camera parameters set in this embodiment are: sensitivity (ISO 100), shutter speed (1 / 1000s), aperture (f / 2.0), and exposure compensation (0.0). The focus mode is set to single autofocus (AFS), and the automatic exposure (AEL) is locked to maintain parameter consistency.

[0200] (3) UAV data collection;

[0201] Start the DJI Mavic 3 Multispectral Edition drone and adjust the camera parameters, including ISO, shutter speed, aperture, exposure compensation, auto focus, and auto exposure lock, to ensure that the camera parameters are consistent for each data collection. The specific parameter settings are shown in Table 3.

[0202] Table 3 Summary of camera shooting parameter settings

[0203] Camera parameters Setting value Image acquisition Visible light image (RGB) & multispectral image (MS) Sensitivity (ISO) 100 Shutter Speed 1 / 1000s Aperture 2.0 Exposure Compensation (ExposureValue) 0.0 Focus Mode (AutoFocus) One-shot autofocus (AFS) AutoExposure Lock locking

[0204] Control the drone to hover 2 meters above the calibration plate and ensure that the calibration plate is in the center of the image;

[0205] Capture multispectral images, generating six images each time: one RGB visible light image, four single-band images, and one four-band composite image; the four single-band images include a green light image, a red light image, a red-edge image, and a near-infrared image (i.e., four-band images of G, R, RE, and NIR);

[0206] Time series acquisition: Repeated shooting at different times of the day (e.g., shooting every 2 hours from 10:00 a.m. to 16:00 p.m.) to cover different solar altitude angles and light intensities;

[0207] The unit of spectral irradiance is W / m 2 / nm, used to describe the radiation power received per unit area and per unit wavelength. The specific breakdown is as follows: W (Watt): unit of power, representing energy per unit time (1 watt = 1 joule / second). 2 (square meter): unit of area, indicating the area covered by the radiant energy. nm (nanometer): unit of wavelength (1 nanometer = 10 -9 The unit quantifies the radiant power received per square meter within a specific wavelength interval (e.g., 1 nanometer width).

[0208] (IV) Image preprocessing (see Figure 8 ):

[0209] (1) Data classification and storage: Classify and store images by different bands (G, R, RE, NIR), extract metadata (EXIF, XMP) to record shooting time, light intensity sensor data, and camera parameters;

[0210] EXIF is primarily used to store camera parameters such as exposure time, aperture value, ISO, focal length, and capture time. For the DJI Mavic 3 Multispectral Edition, EXIF also includes key camera parameters such as sensor gain and black level.

[0211] (2) DN value correction: The digital quantization value is the DN value. According to the black level value (BlackLevel) and gain value (Sensor Gain) in each image metadata, the DN value is subjected to black level correction and gain compensation, thereby eliminating the sensor dark current noise and electronic amplification noise;

[0212] The black level correction formula is: DN corrected =DN raw-DN black DN corrected Represents the DN value after black level correction, DN raw Represents the original digital quantization value (Digital Number), which indicates the DN signal value directly output by the sensor under lighting conditions, with a range of 0-65535 (16-bit camera) or 0-255 (8-bit camera); DN black Represents the black level value, that is, the reference output value of the sensor when there is no light (such as when the lens cover is closed), which is used to characterize the dark current noise of the sensor;

[0213] The specific formula for gain compensation is: DN calibrated =DN corrected / Gain; Gain is the gain coefficient, dimensionless, indicating the sensor's amplification of the signal. DN calibrated Represents the DN value after gain compensation;

[0214] Gain is a parameter of the sensor's electronic amplification circuitry that adjusts signal strength (e.g., ISO 100 corresponds to a gain of 1, ISO 200 corresponds to a gain of 2). The value of Gain is preset by the camera manufacturer or determined experimentally (e.g., photographing a target with known reflectivity and inferring the gain value).

[0215] (3) Distortion correction: Based on the camera calibration parameters (such as lens distortion coefficient and vignetting compensation coefficient), use MATLAB or Python OpenCV tools to perform geometric correction to eliminate image edge distortion;

[0216] The geometric distortion of the corrected image in the edge area is significantly reduced, the image quality is significantly improved, and an image dataset is obtained for further experiments.

[0217] (5) Establishing radiation calibration function:

[0218] (1) Extract the DN value of the calibration plate area;

[0219] In each single-band image, (corresponding to the three calibration plates of black, gray and white, i.e. DN_black, DN_grey and DN_white) the average DN value is calculated;

[0220] (2) Fitting the reflectance-DN relationship model;

[0221] Data pairing is done by associating the known reflectivity of the calibration plate with the corresponding DN value; the known reflectivity of the calibration plate is shown in Table 4.

[0222] Table 4 is a summary table of the standard reflectivity of black, gray and white radiation calibration plates in different bands of light waves.

[0223] Wavelength range 25% blackboard reflectivity 50% gray card reflectivity 75% whiteboard reflectivity Blue (B) 450nm 0.244 0.484 0.739 Green (G) 560nm 0.252 0.498 0.760 Red (R) 650nm 0.255 0.502 0.759 Red edge (RE) 730nm 0.250 0.498 0.752 Near infrared (NIR) 840nm 0.251 0.499 0.750 Near infrared (NIR) 960nm 0.250 0.499 0.744

[0224] The linear function, exponential function and power function are used to fit the DN value and reflectivity. The mathematical expression is:

[0225] The linear function is: R(γ)=G γ ×DN”(γ)+B γ ;

[0226] The exponential function is:

[0227] The power function is:

[0228] In the formula, γ represents one of the five bands, R(γ) is the reflectivity of the corresponding radiation calibration plate in the γ band image, and G γ and B γ is the parameter value, DN”(γ) is the average DN value of the corresponding radiation calibration plate in the corrected γ band image;

[0229] To establish a quantitative relationship between the DN value and reflectivity of the image collected by the spectral camera, the average DN value and reflectivity data of the radiation calibration plate were imported into the Curve Fitting Tool of MATLAB, and linear function, exponential function and power function fitting analysis were performed. In order to evaluate the accuracy and reliability of different fitting models, four key statistical parameters were used to analyze the fitting effect: the sum of squared errors (SSE), the coefficient of determination (R 2 ), Adjusted R 2 ) and Root Mean Squared Error (RMSE);

[0230] First, the residual sum of squares (SSE) is used to measure the deviation between the fitted model and the observed data. Its calculation formula is:

[0231] Secondly, the coefficient of determination (R 2 ) is used to evaluate the ability of the fitted model to explain the variability of the data, and its calculation formula is:

[0232]

[0233] in, is the total sum of squares, is the mean of the observations. R 2 The value range of is [0,1]. The closer it is to 1, the stronger the model's ability to explain the data.

[0234] However, R 2As the number of independent variables in the model increases, it may lead to overfitting problems. Therefore, the Adjusted R 2 ), which is calculated as follows:

[0235]

[0236] Where p is the number of independent variables in the model. 2 It can more objectively evaluate the fitting effect of the model, especially in multiple regression analysis.

[0237] Finally, the root mean square error (RMSE) is used to measure the prediction error of the fitted model, and its calculation formula is:

[0238]

[0239] The smaller the RMSE value, the higher the prediction accuracy of the model.

[0240] By comparing the SSE and R 2 、Adjusted R 2 and RMSE values can be used to comprehensively evaluate the model fitting effect and select the optimal fitting function for subsequent radiation calibration analysis.

[0241] Taking the green (G) band radiation calibration fitting function as an example, linear function, exponential function and power function are fitted, and the fitting functions obtained are shown in Table 5. Subsequently, a scatter plot of reflectivity and average DN value is drawn, and the curves of the three fitting functions are shown in the figure. The results are shown in Table 5. Figure 9 shown.

[0242] Table 5 Green (G) band radiation calibration fitting function

[0243]

[0244] According to the results of the fitting function summary table and the schematic diagram, the following conclusions can be drawn: Among the three fitting models of linear function, exponential function and power function, the exponential function shows the best fitting effect. The SSE and RMSE of the exponential function are significantly lower than those of the linear function and power function, indicating that it has the smallest prediction error. In addition, the R 2 and Adjusted R 2 The exponential function is better than the other two models, indicating that the exponential function can better explain the variability between the average DN value and reflectance. In contrast, the linear function and power function have a slightly worse fitting effect, although their R 2The value is also high, but the SSE and RMSE values are large, indicating that its prediction accuracy is relatively low. In summary, the exponential function is the most suitable model to describe the relationship between the average DN value and reflectivity, providing a reliable theoretical basis for subsequent radiation calibration and quantitative analysis.

[0245] According to the direct radiometric calibration method for the green (G) band camera mentioned above, radiometric calibration was performed on the near-infrared (NIR), red (R), and red-edge (RE) band cameras in turn. The resulting fitting function summary is shown in Table VI.

[0246] Table 6 Three-band radiation calibration fitting function

[0247]

[0248]

[0249] According to the data summarized in the table above, for the near infrared (NIR) band, red light (R) band, and red edge (RE) band, the Adjusted R 2 The values are the highest and the RMSE are the lowest, which shows that the power function model has higher accuracy and reliability in describing the spectral data of these bands. In addition, the linear function and exponential function models also provide good fitting effects, but their Adjusted R 2 The value and RMSE are slightly inferior to those of the power function model. These results show that the power function model is the preferred method for analyzing and fitting the spectral data in these bands.

[0250] To further evaluate the accuracy of the three fitting functions, the present invention calculated the percentage error between the calculated values and the true values. The experimentally measured average DN values were substituted into the fitting equations for the linear function, exponential function, and power function, respectively, to obtain the corresponding calculated reflectance values. The percentage error between the calculated value and the true value was calculated using the formula: Percentage Error = |(Calculated Value - True Value) / True Value | × 100%. The average was then calculated, and the error summary is shown in Table 7. By analyzing the distribution of the percentage error, the fitting accuracy of each fitting function can be further quantified, providing a more reliable basis for selecting the optimal fitting model.

[0251] Table 7 Summary of average percentage errors of camera calibration functions for four bands

[0252]

[0253] A comprehensive analysis of the average percentage errors of the four bands of green light (G), near infrared (NIR), red light (R) and red edge (RE) was conducted. It was found that the overall average percentage errors of the green light (G) band and the red edge (RE) band were the smallest, at 7.47% and 8.96% respectively. Further analysis of the fitting functions of these two bands showed that in the green light (G) band, the exponential function provided the best fitting effect, while in the red edge (RE) band, the power function showed the smallest error. Therefore, this patent adopts these two bands, and uses the exponential function and the power function respectively to perform radiometric calibration between the digital quantization (DN) value and reflectivity of the image captured by the spectral camera.

[0254] Through fitting and analysis, the final radiation calibration function is determined as follows:

[0255] The green light image is matched with a first radiometric calibration function;

[0256] The red edge image is matched with the second radiometric calibration function;

[0257] The first radiation calibration function is: RG = 0.1235*exp(3.043e-05*x); wherein RG is the normalized reflectance of the green light band, and x in this formula is the average DN value of the specific green light image after DN value correction, that is, the DN1.

[0258] The second radiation calibration function is: RRE=5.67e-07*x^1.289; where RRE is the normalized reflectance of the red light band, and x in this formula is the average DN value of the specific red edge image after DN value correction, that is, the DN1.

[0259] The third sub-step of the first step is specifically:

[0260] [min1, max1] is called the first sediment concentration interval, and (max1, max2] is called the second sediment concentration interval. After fitting and evaluation, the final benchmark functions include the green light benchmark function and the red edge benchmark function.

[0261] The green light reference function is: for the first sediment concentration interval, f(x) = 1.528e+17*x^19.38; for the second sediment concentration interval, f(x) = 1.031e+04*x-1551;

[0262] The red edge benchmark function is: for the first sediment concentration interval, f(x) = 2599*x^1.595-1.697; for the second sediment concentration interval, f(x) = 3400*x^1.202-78.31;

[0263] Wherein, f(x) is the sediment content of the water body; x in the first benchmark function is the standardized reflectance of the specific green light image calculated by the first radiation calibration function (i.e., RG value), and x in the first benchmark function is the standardized reflectance of the specific red edge image calculated by the second radiation calibration function (i.e., RRE value).

[0264] After experiments (please refer to the substantive examination reference materials submitted with this application), the green light (G) and red edge (RE) bands exhibit complementary characteristics in different sand content ranges. In sand content range 1, the reflectivity of the green light (G) band is stable at 12.8817% to 15.1959%, while the red edge (RE) band reflectivity has a wider range, ranging from 0.6249% to 5.0721%, but is generally low, which increases the difficulty of monitoring. In sand content range 2, the reflectivity of the red edge (RE) band is significantly increased to 5.0721% to 16.1627%. Not only is the range wider, but the difference is smaller compared to the range of 15.1959% to 18.0259% of the green light (G) band, which improves the feasibility of monitoring.

[0265] min1=0,max1=20,max2=300 (the specific values are in accordance with the characteristics of the Yellow River and are applicable to the Yellow River), and the units are kilograms per cubic meter; [15kg / m 3 ,25kg / m 3 ] is defined as the sediment content transition interval; in the sediment content transition interval, when calculating the sediment content for a specific water body photo, the weighted average of the green light benchmark function and the red edge benchmark function (the mean of the sum of the two) is used as the final water body sediment content; when it is lower than the sediment content transition interval, only the green light benchmark function is used to calculate the water body sediment content; when it is higher than the sediment content transition interval, only the red edge benchmark function is used to calculate the water body sediment content.

[0266] The weighted average method can avoid model switching errors and also make the calculation results transition smoothly.

[0267] When establishing the benchmark function, an outdoor experiment was first conducted. The sand used in the experiment was taken from natural sand samples on both sides of the waters near the Huayuankou hydrological station of the Yellow River. In order to fully reflect the characteristics of the Yellow River sediment, three sand samples with different particle sizes and dryness and humidity were collected at Huayuankou of the Yellow River at 5 meters, 30 meters, and 50 meters away from the Yellow River channel, such as Figure 2 shown.

[0268] Sand collection point 1: Located 50 meters from the Yellow River channel, approximately 75 kg of fine-grained dry sand was collected;

[0269] Sand collection point 2: Located 30 meters from the Yellow River channel, approximately 70 kg of coarse dry sand was collected;

[0270] Sand collection point 3: Located 5 meters away from the Yellow River channel, about 70 kilograms of coarse-grained wet sand was collected.

[0271] The collected sand samples are dried and then thoroughly mixed to ensure that the sand samples can represent the physical properties of the Yellow River sediment in the area as much as possible.

[0272] The container used in this experiment is a cylindrical plastic basin with a diameter of 80 cm and a height of 30 cm, closed at the bottom and open at the top. To suppress the reflection of light from the bottom and side walls of the container, the entire interior of the container is covered with a black anti-seepage geomembrane. During the experiment, clean water is first poured into the container to a depth of 20 cm. At this time, the water volume is 0.1 m 3 .

[0273] According to literature research, the sediment content in the waters near the Huayuankou hydrological station of the Yellow River is stable at [0 kg / m 3 ,20kg / m 3 ] range, but during the period of water and sand regulation, the sediment content in this area will change significantly and exceed the normal range. Therefore, this paper further prepares (20kg / m 3 ,300kg / m 3 Experiments were conducted on water bodies with a range of sediment concentrations to construct a more complete benchmark function between sediment-laden water and light reflectivity. The experimental sediment concentrations are shown in Table 8.

[0274] Table 8 Experimental sand content preparation table

[0275] Number of groups <![CDATA[Sediment concentration (kg / m 3 )]]> Amount of sand added (g) Incremental amount (g) 1 0 0 0 2 2 201.21 201.21 3 4 402.72 201.51 4 6 604.52 201.81 5 8 806.63 202.10 6 10 1009.03 202.40 7 12 1211.74 202.70 8 14 1414.74 203.00 9 16 1618.05 203.31 10 18 1821.66 203.61 11 20 2025.57 203.91 12 60 6168.43 4142.86 13 100 10438.27 4269.84 14 140 14841.03 4402.76 15 180 19383.01 4541.98 16 220 24070.91 4687.90 17 260 28911.88 4840.98 18 300 33913.56 5001.67

[0276] After the container is filled with water, start the drone and adjust the drone's gimbal to allow the multispectral camera to shoot vertically at the height of 1m, 2m, 3m, 4m, and 5m above the water surface. The experimental overview is shown in the figure below. Figure 3 The drone camera parameters were set to the same settings as in the previous camera radiometric calibration experiment (see Table 3 for specific parameters). Each capture generated six images, including a standard RGB image, a four-band spectral composite image, and four single-spectral images.

[0277] After the initial shooting is completed, dry sand is gradually added to the clean water according to the sediment content gradient listed in Table 8 to prepare water bodies with different sediment contents. The water body is fully stirred using a stirrer to keep the sediment in the water body in a suspended state and prevent it from settling. In this way, a sandy water body with a fixed sediment content and uniform sand distribution is prepared. After each preparation is completed, the drone is launched again and multi-height data is collected according to fixed parameters. The image data example is as follows Figure 4As shown. Repeat the above steps until the data collection of all sediment gradients is completed. Since this experiment aims to simulate the sediment content monitoring conditions of natural rivers, all experiments are carried out in outdoor environments. Each experiment is strictly carried out in accordance with the sediment content gradient preparation and data collection process to construct an accurate and complete benchmark function of the surface sediment content of sandy water bodies and the reflectivity of the surface water body to light waves. In the experiment, a black anti-seepage geomembrane was laid on the bottom of the container to effectively absorb the transmitted light and prevent the secondary reflection of the light at the bottom of the container from affecting the measurement of the surface reflectivity. Ensure that the water depth meets the requirement that the multispectral camera cannot capture the information of the bottom of the container (when shooting a river, the riverbed cannot be photographed), meet the actual shooting conditions during actual measurement, and avoid the image of the bottom of the container interfering with the experiment, to ensure that the subsequent sediment content data represent the sediment content within the depth range.

[0278] After the experimental data collection is completed, the image data taken by the multispectral camera is first classified according to the band, which is divided into green light (G) and red edge (RE) band images, so that the images of different bands can be processed independently. Taking the green light (G) band image data as an example, the data set contains 18 sediment content gradients (0 kg / m 3 ~300kg / m 3 ) at five flight altitudes (1m, 2m, 3m, 4m, and 5m). Following the image preprocessing method described above, all image data used to construct the "reflectivity-sediment content" benchmark function were subjected to DN value correction and distortion correction to obtain the preprocessed benchmark function image data. The DN value at this point is recorded as DN'(G).

[0279] The preprocessed image data is then used to read its DN value. Specifically, water areas are manually demarcated in each drone-collected image, removing interference from objects such as waves and floating objects. The average DN value of the demarcated water area is then calculated and recorded as "DN" (G).

[0280] Comparing the DN values before and after data preprocessing reveals a significant downward shift in the average value of the preprocessed data compared to the original data. This phenomenon is primarily due to the black level correction and sensor gain elimination performed on the original image data during preprocessing, which effectively reduces system errors. Furthermore, the trends of the original and preprocessed data are largely consistent.

[0281] Specifically, when the sand content is 0kg / m 3 ~2kg / m 3The DN value increases fastest when the test container is filled with water. This is because the bottom and sides of the test container are painted black, which has a strong ability to absorb light. When only clean water is injected, the water has the strongest light transmittance, and only a very small amount of light is reflected, resulting in a darker image captured by the camera. As sediment is gradually added to the clean water, the water becomes turbid, the light transmittance decreases, the reflectivity increases, and the image captured by the camera becomes brighter, so the DN value increases significantly. In the sand content range of 2kg / m 3 ~20kg / m 3 When the sand content is within the range of 20kg / m 3 ~300kg / m 3 When the altitude is 1m, the growth rate of the DN value is further accelerated. This is due to the influence of the increased sediment gradient, which leads to a significant increase in the growth rate of the DN value and the slope of the summary linear image. In addition, when analyzing the image data taken at different flight altitudes, it was found that the DN value of the image data taken at a flight altitude of 1m was significantly higher than that at other altitudes. This shows that the data at this altitude has a large abnormality. When the flight altitude is 2m, 3m, 4m, and 5m, the DN value of the image data taken at the four altitudes is less different, and the fluctuation range is also smaller, which shows that the data have a high consistency.

[0282] Based on the above analysis, the image data captured at a flight altitude of 1 m was excluded from subsequent analyses, and only the data from altitudes of 2 m, 3 m, 4 m, and 5 m were used. Furthermore, because the DN values at these four altitudes differed slightly, their mean was used as the DN value for establishing the "reflectivity-sediment content" benchmark function in subsequent studies.

[0283] In order to further improve the accuracy of the benchmark function, the sediment content gradient is divided into two intervals for fitting. Interval 1: [0 kg / m 3 ,20kg / m 3 ]; Interval 2: (20kg / m 3 ,300kg / m 3 ]. Through piecewise fitting, the relationship between DN value and sediment content in different sediment content ranges can be better reflected, thereby improving the applicability and accuracy of the benchmark function.

[0284] The red edge (RE) band spectral image data was analyzed in the same way, and it was found that it had the same rules as the green light (G) band spectral image data. Therefore, the red edge (RE) band spectral image data was processed in the same way as the green light (G) band spectral image data.

[0285] Establishment and analysis of two-band benchmark functions

[0286] According to the optimal radiometric calibration function constructed for the green (G) band camera mentioned above, the digital quantization (DN) value of the image data is converted into reflectivity. In order to establish a quantitative relationship between the reflectivity of the surface layer of the sand-laden water body to the green (G) band light wave and the sand content of the water body surface, a scatter plot corresponding to the reflectivity and sand content is drawn, as shown in the figure below. Figure 5 shown.

[0287] The scattered data were imported into the Curve Fitting Tool of MATLAB, and various function fitting analyses were performed, including linear function, exponential function, and power function. In order to evaluate the accuracy and reliability of different fitting models, four key statistical parameters were also used to analyze the fitting effect: residual sum of squares (SSE), coefficient of determination (R 2 ), Adjusted R 2 ) and root mean square error (RMSE).

[0288] Regarding the fitting function relationship between the reflectivity of the surface layer of the sand-laden water body to the green light (G) band light wave and the sand content of the water body surface layer, the fitting function results obtained in the sand content interval 1 and the sand content interval 2 are shown in detail in Table 9.

[0289] Table 9 Summary of fitting functions of green light (G) band reflectivity and sediment content

[0290]

[0291]

[0292] Goodness-of-fit analysis revealed significant differences between the different function models in the two sediment concentration intervals. In sediment concentration interval 1, a clear gradient in the fitting performance of the three function models was observed: the power function exhibited the best fitting performance, with the smallest sum of squared residuals and the highest coefficient of determination, while the exponential and linear functions exhibited larger fitting errors. Based on this, the power function was selected as the fitting model for green band reflectance and sediment concentration in interval 1.

[0293] In sediment concentration interval 2, both the linear function and the combined exponential function showed excellent fitting performance. Among them, the SSE of the linear function was 245.3, and the adjusted R 2 The p-value reached 0.9957, indicating high fitting accuracy and explanatory power. The combined exponential function provided a better fit, but its model complexity was relatively high. In contrast, the power function and single exponential function provided weaker fits. Considering model accuracy, computational efficiency, and practicality, the linear function was selected as the fitting model for interval 2.

[0294] In summary, the quantitative relationship between the reflectivity of the surface layer of the sand-laden water body to the green light (G) band light wave and the sediment content of the water body surface layer is described by a power function in sediment content interval 1 (i.e., the first sediment content interval), and by a linear function in sediment content interval 2 (i.e., the second sediment content interval). These two fitting function images are shown in Figure 1. Figure 6 shown.

[0295] The same method was used to establish a quantitative relationship between the reflectivity of the surface layer of sand-laden water to red-edge (RE) band light and the sediment content of the surface layer. The fitting functions for intervals 1 and 2 are summarized in Table 10.

[0296] Table 10 Summary of fitting functions of red edge (RE) band reflectivity and sediment content

[0297]

[0298]

[0299] According to the comparative analysis of the fitting results, the power function after translation shows the best performance in both intervals. In interval 1, the SSE of the power function after translation is 9.159, R 2 The power function achieved a fitting accuracy of 0.9792, demonstrating excellent fitting accuracy and error control, indicating its ability to well characterize the nonlinear relationship between reflectance and sediment concentration. In interval 2, the power function maintained optimal fitting performance after translation, significantly outperforming other models in fitting accuracy and stability, and applicable to a wider range of sediment concentrations.

[0300] In summary, the quantitative relationship between the reflectivity of the surface layer of the sediment-laden water body to the red edge (RE) band light wave and the sediment content of the water body surface layer is described uniformly by using the shifted power function in sediment content interval 1 and sediment content interval 2, and is selected as the benchmark function. For the two fitting function images (i.e., red edge benchmark functions) of the two sediment content intervals, Figure 7 shown.

[0301] Summarizing the above analysis, we obtain the red edge benchmark function and the green light benchmark function, forming a complete benchmark function.

[0302] The above embodiments are only used to illustrate rather than limit the technical solutions of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the present invention can still be modified or replaced by equivalents. Any modification or partial replacement that does not depart from the spirit and scope of the present invention should be included in the scope of the claims of the present invention.

Claims

1. The method for measuring sediment content in water bodies based on drones is characterized by Follow these steps: The first is to establish an inference model from the digital quantification value of the target water body image to the sediment content of the target water body; The second is that the drone flies along the target water body and collects images of different areas of the target water body and transmits the images back to a remote computer; The third is to manually demarcate the water area in each image collected by the drone; Fourth, the computer calculates the average value DN1 of the digital quantization value of each pixel for the marked water area in each image; The fifth is that through the inference model established in the first step, the computer obtains the sand content corresponding to each image collected by the drone based on the DN1 value corresponding to each image collected by the drone.

2. The method for measuring sediment content in water bodies based on an unmanned aerial vehicle according to claim 1, characterized in that: The first step includes the following sub-steps: First, for the target water body, confirm its daily sediment content range [min1, max1] and the sediment content range during water and sediment regulation (max1, max2] by measuring or querying historical data, where min1, max1, and max2 are all in kilograms per cubic meter; The second is to establish a radiometric calibration function to determine the reflectivity of water images collected by UAVs; The third is to establish a benchmark function to establish the functional relationship between reflectivity and sediment content of the target water body.

3. The method for measuring sediment concentration in water bodies based on an unmanned aerial vehicle according to claim 2, characterized in that: In the first sub-step of the first step, min1=0, max1=20, and max2=300.

4. The method for measuring sediment concentration in water bodies based on an unmanned aerial vehicle according to claim 2, characterized in that: The second sub-step of the first step is specifically: (1) Hardware preparation: A drone equipped with a multispectral camera was used. A light intensity sensor was integrated on the top of the drone to record the spectral irradiance of each band during shooting. Three calibration plates (15 cm × 15 cm in size) were used: black, gray, and white. The standard reflectance of each band was pre-measured by a hyperspectral instrument. (2) Environmental requirements: Good outdoor visibility to avoid shadows or debris covering the calibration plate; the drone's flight altitude is fixed at 2 meters to ensure that the entire calibration plate is fully captured by the multispectral camera; (3) Data collection process: (1) Calibration board layout: Place three calibration boards (black, gray, and white) horizontally on an open ground. Keep the surface clean and free of debris to avoid reflection interference. Use a level to ensure that the calibration boards are parallel to the ground to reduce geometric distortion. (2) Set the camera parameters; (3) Collect drone data; Control the drone to hover 2 meters above the calibration plate and ensure that the calibration plate is in the center of the image; Capture multispectral images, generating six images per capture: one RGB visible light image, four single-band images, and one four-band composite image; the four single-band images include a green light image, a red light image, a red-edge image, and a near-infrared image; Time series acquisition: repeated shooting at different times of the day to cover different sun altitudes and light intensities; (IV) Image preprocessing: (1) Data classification and storage: Classify and store images by different bands, extract metadata to record shooting time, light intensity sensor data, and camera parameters; (2) DN value correction: The digital quantization value is the DN value. According to the black level value and gain value in each image metadata, the DN value is subjected to black level correction and gain compensation, thereby eliminating the sensor dark current noise and electronic amplification noise; (3) Distortion correction: Based on the camera calibration parameters, use MATLAB or Python OpenCV tools to perform geometric correction to eliminate image edge deformation; (5) Establishing radiation calibration function: (1) Extract the DN value of the calibration plate area; In each single-band image, the average DN value was calculated; (2) Fitting the reflectance-DN relationship model; Data matching, associating the known reflectivity of the calibration plate with the corresponding DN value; The linear function, exponential function and power function are used to fit the DN value and reflectivity. The mathematical expression is: The linear function is: R(γ)=G γ ×DN”(γ)+B γ ; The exponential function is: The power function is: In the formula, γ represents one of the five bands, R(γ) is the reflectivity of the corresponding radiation calibration plate in the γ band image, and G γ and B γ is the parameter value, DN”(γ) is the average DN value of the corresponding radiation calibration plate in the corrected γ band image; To establish a quantitative relationship between the DN value and reflectivity of images captured by a spectral camera, the average DN value and reflectivity data of the radiometric calibration plate were imported into MATLAB's Curve Fitting Tool for linear, exponential, and power function fitting analysis. To evaluate the accuracy and reliability of different fitting models, four key statistical parameters were used to analyze the fitting results: residual sum of squares, coefficient of determination, adjusted coefficient of determination, and root mean square error. Through fitting and analysis, the final radiation calibration function is determined as follows: The green light image is matched with a first radiometric calibration function; The red edge image is matched with the second radiometric calibration function; The first radiation calibration function is: RG = 0.1235*exp(3.043e-05*x); where RG is the normalized reflectance of the green light band, and x in this formula is the average DN value of the specific green light image after DN value correction; The second radiation calibration function is: RRE = 5.67e-07*x^1.289; where RRE is the normalized reflectance of the red light band, and x in this formula is the average DN value of the specific red edge image after DN value correction.

5. The method for measuring sediment content in water bodies based on an unmanned aerial vehicle according to claim 2, characterized in that: The third sub-step of the first step is specifically: [min1, max1] is called the first sediment concentration interval, and (max1, max2] is called the second sediment concentration interval. After fitting and evaluation, the final benchmark functions include the green light benchmark function and the red edge benchmark function. The green light reference function is: for the first sediment concentration interval, f(x) = 1.528e+17*x^19.38; for the second sediment concentration interval, f(x) = 1.031e+04*x-1551; The red edge benchmark function is: for the first sediment concentration interval, f(x) = 2599*x^1.595-1.697; for the second sediment concentration interval, f(x) = 3400*x^1.202-78.31; Wherein, f(x) is the sediment content of the water body; x in the first benchmark function is the standardized reflectance of the specific green light image calculated by the first radiation calibration function, and x in the first benchmark function is the standardized reflectance of the specific red edge image calculated by the second radiation calibration function.

6. The method for measuring sediment concentration in water bodies based on an unmanned aerial vehicle according to claim 5, characterized in that: min1=0,max1=20,max2=300,the unit is kilograms per cubic meter; 3 ,25kg / m 3 ] is defined as the sediment content transition interval; in the sediment content transition interval, when calculating the sediment content for a specific water body photo, the weighted average of the green light benchmark function and the red edge benchmark function is used as the final water body sediment content; when it is lower than the sediment content transition interval, only the green light benchmark function is used to calculate the water body sediment content; when it is higher than the sediment content transition interval, only the red edge benchmark function is used to calculate the water body sediment content.

Images