Lake surface fluctuation monitoring method based on unmanned aerial vehicle

Abstract

The present application relates to the technical field of water conservancy and water environment monitoring, and discloses a lake surface fluctuation monitoring method based on a UAV, comprising: S1, periodically cruising the monitored water area according to a preset route, obtaining multi-temporal multispectral images and visible light images, and extracting an actual water surface area; S2, based on real-time hydrological and water level data of the monitored water area, establishing a water level-area mathematical relationship model of the monitored water area, and uniformly calibrating the actual water surface area to a standardized area under a preset reference water level; S3, after obtaining the water surface area change rate and change polygon data of the monitored water area in multiple years and multiple periods, using a time series analysis model to predict the future area change trend of the water area, and obtaining a prediction coefficient; S4, identifying a risk area according to the prediction coefficient, monitoring the risk area, and generating a dynamic polygon of the risk area and a scoring level; and S5, based on the dynamic polygon, publicly disclosing risk information in three dimensions.

Description

Technical Field

[0001] This invention relates to the field of water conservancy engineering and water environment monitoring technology, specifically to a method for monitoring lake surface fluctuations based on unmanned aerial vehicles (UAVs). Background Technology

[0002] Lakes and reservoirs are among the most important freshwater resources on Earth, playing an irreplaceable role in regulating the terrestrial water cycle, maintaining the balance of wetland ecosystems, and serving as global carbon sinks. However, since the 1980s, under the dual pressures of intensified human activities and abnormal climate, water bodies in different regions of the world have shown varying degrees of shrinkage or expansion. Taking the Inner Mongolia Plateau in China as an example, under the combined effects of a warming and aridifying climate and increased human activity, lakes have experienced significant shrinkage and degradation, leading to a series of ecological problems such as reduced biodiversity, water salinization, and land degradation. Furthermore, frequent human encroachment on water bodies has significantly impacted their functions, including flood control and water supply. Therefore, accurate and dynamic monitoring and assessment of water area are of paramount importance for ensuring flood control, water supply, and ecological security.

[0003] Currently, water area monitoring mainly relies on manual patrols or satellite remote sensing. Manual patrols are limited by terrain and funding, resulting in fewer patrols and fixed routes, leading to incomplete patrols. Optical satellite remote sensing is greatly affected by cloud and rain weather, and in areas such as the Yangtze-Huaihe River Basin, the annual effective data acquisition rate is less than half, resulting in untimely detection of water area encroachment and high cost and difficulty in handling it. Existing methods, such as ground-based automatic detection using buoy monitoring stations and shore-based cameras, can achieve continuous, fixed-point observations unaffected by weather and monitor multiple water quality indicators. However, limitations include high cost, limited installation locations, fixed camera positions, and a restricted field of view, failing to cover the spatial distribution information of large water areas. Furthermore, with the continuous maturation of drone technology, drones are being widely used in aerial surveying, water conservancy, and environmental industries, unaffected by cloud cover or other environmental factors. Multispectral remote sensing technology has been successfully applied to the quantitative inversion of water quality parameters, and time series analysis models such as LSTM have demonstrated significant value in predicting water evolution. However, some shortcomings still exist. First, area and water quality monitoring are disconnected; existing drone applications focus on a single objective—either monitoring changes in water area or retrieving water quality parameters, lacking a comprehensive monitoring and assessment method that considers both in a unified manner. Second, there is a lack of water level correction mechanism; the water area is significantly affected by seasonal water level fluctuations, but existing methods mostly directly compare area data from different periods without establishing a quantitative relationship model between water level and area for correction, resulting in monitoring results that cannot truly reflect area changes caused by human factors. Third, it lacks trend prediction and risk warning functions; existing methods mainly rely on "current situation assessment" or "historical comparison", which are ex-post supervision and have not yet formed an active early warning capability based on time series prediction models. Summary of the Invention

[0004] The purpose of this invention is to provide a method for monitoring lake surface fluctuations based on unmanned aerial vehicles (UAVs), and to solve the following technical problems: How to achieve all-weather, high-precision, and intelligent dynamic monitoring and trend prediction of changes in water area?

[0005] The objective of this invention can be achieved through the following technical solutions: A method for monitoring lake surface fluctuations based on unmanned aerial vehicles (UAVs), the method includes: S1. Conduct periodic patrols of the monitored water area according to the preset route to acquire multispectral and visible light images at multiple time intervals and extract the actual water surface area. S2. Based on real-time hydrological and water level data of the monitored water area, establish a mathematical relationship model between water level and area of ​​the monitored water area, and uniformly calibrate the actual water surface area to the standardized area under the preset benchmark water level. S3. After obtaining the water surface area change rate and change patch data of the monitored water area over many years and multiple periods, the time series analysis model is used to predict the future area change trend of the water area and obtain the prediction coefficient. S4. Identify risk areas based on prediction coefficients, monitor risk areas, and generate dynamic maps and rating levels for risk areas; S5, and displays risk information in three dimensions based on dynamic patterns.

[0006] Preferably, the method for extracting the water surface area is as follows: Based on visible light imagery, the normalized water index is calculated, the instantaneous water boundary line of the monitored lake is extracted, the corresponding instantaneous water area is calculated, and the data is compared with historical time series data to obtain the lake surface area fluctuation characteristics.

[0007] Preferably, the normalized water index is calculated as follows:

[0008] in, The reflectivity in the green light band, This refers to the reflectivity in the near-infrared band.

[0009] Preferably, the specific method for determining the instantaneous water area based on the normalized water index is as follows: Determining the adaptive segmentation threshold using the maximum entropy method When the NDWI value of any pixel in a visible light image is greater than the segmentation threshold When this happens, the pixel is identified as a water area pixel; the total number of all water area pixels is counted. Combined with the ground resolution of visible light imagery Calculate the instantaneous water area Calculation formula:

[0010] in, The unit is square meters.

[0011] Preferably, the method for constructing the water level-area mathematical relationship model of the monitored water area in step S2 includes: S21. Collect water level observation data of the monitored water area at multiple historical periods. and the corresponding measured water area This constitutes a water level-area sample dataset. ; S22. Use nonlinear regression to fit the sample dataset and establish the water level-area function relationship:

[0012] in, For real-time hydrological water level data, For the corresponding water area, , , Parameters for model fitting; S23. Based on the water level-area function relationship, the actual water surface area... Uniform calibration to the preset reference water level Standardized area The calibration formula is:

[0013] in, This refers to the real-time water level at the moment of aerial photography. This refers to the area corresponding to the current water level, calculated based on the water level-area function. This represents the area corresponding to the benchmark water level, calculated based on the water level-area function.

[0014] Preferably, the methods for obtaining the water surface area change rate and change patch data of the monitored water area over many years and multiple periods in step S3 include: S31. Based on the standardized area obtained from multiple consecutive monitoring periods, calculate the... The monitoring cycle is relative to the first Water area change rate over monitoring periods The calculation formula is:

[0015] in, For the first Standardized area for each monitoring cycle for Standardized area for each monitoring cycle; S32. Perform spatial overlay analysis on the standardized areas of adjacent monitoring periods to mark the spatial areas where the water boundary changes, and generate change patch data. The change patch data includes the latitude and longitude information, location information, area information and change nature information of the change patches. S33. The sequence of water area change rates obtained over multiple years and periods. The change patch data serves as the basic input data for the time series analysis model.

[0016] Preferably, in step S3, the time series analysis model is used to predict the future area change trend of the water body, and the prediction coefficients are obtained by means of: S34. Construct an LSTM (Long Short-Term Memory) network model. The LSTM model includes an input layer, an LSTM hidden layer, and an output layer. S35. Analyze the historical time series of water area change rates. The training sample set is constructed by sliding a time window, using the input sequence as input; the input of each training sample is continuous. Change rate data at each historical time step Output label is future Change rate data at each time step ; S36. The LSTM model is trained using the backpropagation algorithm and the adaptive moment estimator optimizer, with the goal of minimizing the mean squared error loss function between the predicted value and the true value, to obtain the trained LSTM prediction model. S37, Move the most recent consecutive The water area change rate data at each time step is input into the trained LSTM prediction model, and the model outputs a predicted area change rate sequence for the next preset time step, which serves as the prediction coefficient.

[0017] Preferably, the prediction coefficient includes the formula for calculating the predicted area change rate:

[0018] in, For the future Predicted rate of change of water area over a time step For time window The historical area change rate input sequence within the region, This is the set of training weight parameters for the LSTM model. Predictive coefficients also include risk level probability vectors The calculation formula is as follows:

[0019] in, For LSTM models in the future The hidden state vector output at each time step. and Here are the weight matrix and biases for the fully connected layer; , , , , These correspond to multiple preset risk levels of probability values, where 0 < < < < < <1, ,and ∈ ; The prediction coefficient also includes the absolute value of the predicted area. The calculation formula is as follows:

[0020] in, This refers to the standardized area for the current assessment period.

[0021] Preferably, the method for identifying risk areas based on prediction coefficients in step S4 includes: Preset risk level thresholds, including high-risk probability thresholds. Medium risk probability threshold and low-risk probability threshold ; When the high-risk level probability in the risk level probability vector Greater than or equal to the high-risk probability threshold At that time, the area was identified as a high-risk area; When the probability of a high-risk level Less than the high-risk probability threshold And the probability of medium risk level Greater than or equal to the medium risk probability threshold At that time, the area was identified as a medium-risk area; When all risk level probabilities in the risk level probability vector are lower than the corresponding threshold, the area is determined to be a low-risk area. The rating levels correspond to the risk levels: high-risk areas correspond to the first rating level, medium-risk areas correspond to the second rating level, and low-risk areas correspond to the third rating level.

[0022] Preferably, after generating the dynamic patches and rating levels of the risk area in step S4, the method further includes a step of generating early warning information: The dynamic patches are compared with pre-stored historical change patches and predicted dynamic patches; the predicted dynamic patches are the expected change areas within the future prediction period generated based on prediction coefficients. Calculate the deviation rate between the area of ​​the dynamic patch and the predicted area of ​​the dynamic patch. The calculation formula is:

[0023] in, This represents the actual area of ​​the dynamic patch. To predict the area of ​​dynamic patches; Preset deviation rate threshold When the deviation rate The absolute value is greater than or equal to the deviation rate threshold. When this occurs, the early warning mechanism is triggered, and early warning information is generated; Early warning information includes the warning level, warning type, location of the risk area, predicted evolution trend, and recommended response measures; the warning level is based on the deviation rate. The magnitude of the deviation rate is divided into several levels, and the larger the absolute value of the deviation rate, the higher the warning level.

[0024] The beneficial effects of this invention are: (1) This invention utilizes a UAV equipped with a multispectral sensor and a high-resolution optical camera to periodically cruise along a preset route, acquiring multispectral and visible light images at multiple time intervals. During image processing, the Normalized Difference Water Index (NDWI) combined with the maximum entropy method is used to adaptively determine the segmentation threshold, which can effectively enhance water body information and suppress interference from vegetation and soil information, achieving accurate identification of water area pixels and thus accurately extracting the instantaneous water area. Based on this, by collecting water level observation data over many years and the corresponding measured water area, a water level-area nonlinear regression mathematical model is constructed. The instantaneous water surface area obtained under different water level conditions is uniformly calibrated to the standardized area under the preset benchmark water level, eliminating the interference of water level fluctuations on area measurement. This makes the monitoring data under different periods and water level conditions have a unified comparable benchmark, significantly improving the accuracy and consistency of long-term monitoring data.

[0025] (2) This invention calculates the water area change rate of the j-th monitoring period relative to the (j-1)-th monitoring period based on the standardized area obtained from multiple consecutive monitoring periods, and performs spatial overlay analysis on the standardized areas of adjacent monitoring periods to mark the spatial areas where the water boundary changes, generating change patch data containing latitude and longitude information, location information, area information and change nature information, providing basic input data for the time series analysis model; on this basis, an LSTM long short-term memory network model including an input layer, an LSTM hidden layer and an output layer is constructed, the historical water area change rate sequence is used as the input sequence, a training sample set is constructed by sliding the time window, and the backpropagation algorithm and adaptive moment estimation optimizer are used for training to obtain a trained LSTM prediction model, the water area change rate data of the most recent m consecutive time steps are input into the model, and the predicted area change rate sequence, risk level probability vector and predicted area absolute value of the future preset time step are output as prediction coefficients. By setting a preset risk level threshold, and comparing the high-risk probability in the risk level probability vector with the high-risk probability threshold, the region is determined to be a high-risk region, a medium-risk region, or a low-risk region, and corresponding dynamic patches and scoring levels are generated, realizing the transformation from passive monitoring to proactive prediction and from qualitative analysis to quantitative assessment.

[0026] (3) After generating dynamic patches and rating levels for risk areas, this invention further compares the dynamic patches with pre-stored historical change patches and predicted dynamic patches generated based on prediction coefficients. It calculates the deviation rate between the area of ​​the dynamic patches and the area of ​​the predicted dynamic patches, and presets a deviation rate threshold. When the absolute value of the deviation rate is greater than or equal to the threshold, an early warning mechanism is triggered, generating early warning information including the early warning level, early warning type, risk area location, predicted evolution trend, and recommended disposal measures. The early warning level is divided into several levels based on the magnitude of the deviation rate; the larger the absolute value of the deviation rate, the higher the early warning level, forming a closed-loop management system from monitoring, prediction, anomaly identification to disposal recommendations. Simultaneously, based on the dynamic patches, risk information is displayed in three dimensions, making the spatial location, evolution trend, rating level, and early warning information of the risk area publicly visible and subject to public supervision. This not only improves the emergency response efficiency of management departments but also enhances the public's right to know and participate in water area risks.

[0027] Of course, any product implementing this invention does not necessarily need to achieve all the advantages described above at the same time. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart illustrating the steps of the UAV-based lake surface fluctuation monitoring method of the present invention. Figure 2 This is a flowchart illustrating the process of obtaining prediction coefficients by using a time series analysis model to predict the future area change trend of water bodies in step S3 of the present invention. Detailed Implementation

[0030] The technical solutions of 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.

[0031] Please see Figure 1 As shown, this invention is a method for monitoring lake surface fluctuations based on unmanned aerial vehicles (UAVs), the method comprising: S1. Based on the multispectral sensor and high-resolution optical camera carried by the UAV, periodically patrol the monitored water area according to the preset route to acquire multispectral and visible light images in multiple time series and extract the actual water surface area. S2. Based on real-time hydrological and water level data of the monitored water area, establish a mathematical relationship model between water level and area of ​​the monitored water area, and uniformly calibrate the actual water surface area to the standardized area under the preset benchmark water level. S3. After obtaining the water surface area change rate and change patch data of the monitored water area over many years and multiple periods, the time series analysis model is used to predict the future area change trend of the water area and obtain the prediction coefficient. S4. Identify risk areas based on prediction coefficients, monitor risk areas, and generate dynamic maps and rating levels for risk areas; S5, and displays risk information in three dimensions based on dynamic patterns.

[0032] In the above technical solution, firstly, step S1, based on the multispectral sensor and high-resolution optical camera carried by the UAV, periodically patrols the monitored water area according to a preset route to acquire multispectral and visible light images of multiple time series and extract the actual water surface area; based on the lake management area and spatial data such as dikes, the UAV aerial photography range is determined, the route is laid along the lake shoreline, the forward coverage extends beyond the operation boundary line by no less than two baselines, and the lateral coverage extends beyond the operation boundary line by no less than 50% of the image frame. Specific route planning parameters include: yaw rate < ±5m, altitude difference < 5m; forward overlap no less than 70%; lateral overlap no less than 55%; flight route altitude 350 meters, survey area 290 square kilometers, and 95 flight sorties; the UAV, equipped with a multispectral sensor and high-resolution optical camera, periodically patrols the lake surface according to the above preset route, with a patrol cycle of once a month, acquiring multispectral and visible light images of multiple time series, with a ground resolution better than 0.3m; For the acquired visible light images, the Normalized Difference Water Index (NDWI) is calculated, which can effectively enhance water information and suppress vegetation and soil information, making water appear as a significantly bright feature in the image. The maximum entropy method is used to determine the adaptive segmentation threshold, and any pixel in the visible light image is identified as a water pixel when its NDWI value is greater than the segmentation threshold.

[0033] Then, in step S2, water level observation data and corresponding measured water area of ​​the lake surface over the past few years are collected to form a water level-area sample dataset for fitting, and a water level-area function relationship is established. The water level-area relationship model of Luoma Lake is obtained by fitting with the least squares method, and the fitting effect is judged according to the model generation result. Based on the above water level-area function relationship, the actual water surface area is uniformly calibrated to the standardized area under the preset benchmark water level. In this embodiment, the normal water storage level of the lake surface is selected as the benchmark water level, and the standardized area is determined after water level calibration. Next, step S3 calculates the water area change rate based on the standardized area obtained from multiple consecutive monitoring periods, performs spatial overlay analysis on the standardized areas of adjacent monitoring periods, marks the spatial areas where the water boundary changes, and generates change patch data. The change patch data includes the latitude and longitude information, location information, area information, and change nature information (increase or decrease in water area) of the change patches. Through spatial analysis, the number of change patches is identified, and patches with increased area are identified (mainly ecological restoration areas), and 2 patches with decreased area are identified (further on-site verification is required). By constructing an LSTM (Long Short-Term Memory) network model, multiple periods of water area change rate sequences are used as training datasets. A sliding time window method is used to construct the training sample set. The input of each training sample is the change rate data of the continuous period, and the output label is the change rate data of the future period. The training set and the test set are divided in an 8:2 ratio. After a certain number of training iterations, the mean squared error and mean absolute percentage error of the model on the test set meet the set requirements, ensuring that the prediction accuracy meets the requirements of engineering applications. In addition, the trained LSTM prediction model is input with the water area change rate data of the continuous period, and the model outputs the predicted area change rate sequence as the prediction coefficients. Furthermore, step S4 determines the probability thresholds for different risk levels based on the preset risk level thresholds, and takes the identified multiple lake surface area reduction patches as examples to calculate the probability of high-risk registration and determine whether they are high-risk or low-risk areas. Based on the dynamic patches of the generated risk areas, high-risk areas are monitored in a focused manner, and the frequency of drone patrols is increased to shorten the cycle length. In addition, by comparing the dynamic patches of high-risk areas with the predicted dynamic patches, the deviation rate between the areas of the dynamic patches and the predicted dynamic patches is calculated. Based on the pre-determined dynamic patch area of ​​high-risk areas, the dynamic patch area and deviation rate are predicted. Based on the predicted deviation rate threshold, the early warning mechanism is triggered, and early warning information is generated. The early warning level is generated according to the deviation rate, and the early warning type is determined based on the early warning level information. The early warning content includes the location of the risk area, the predicted evolution trend, and recommended response measures.

[0034] Finally, the dynamic map features are overlaid on the three-dimensional digital elevation model of the lake surface and visualized using a WebGIS platform. The displayed content includes the spatial location of the risk area, dynamic simulation of its evolution trend, rating level, and early warning information. At the same time, the risk information is made public to the public through government affairs apps and WeChat mini programs, thereby accepting public supervision.

[0035] As one embodiment of the present invention, the method for extracting the water surface area is as follows: Based on visible light imagery, a normalized water index is calculated to extract the instantaneous water boundary of the monitored lake, calculate the corresponding instantaneous water area, and compare it with historical time-series data to obtain the lake surface area fluctuation characteristics. Specifically, the normalized water index is calculated as follows:

[0036] in, The reflectivity in the green light band, The reflectance is in the near-infrared band; when extracting the instantaneous waterline, the maximum entropy method is used to determine the adaptive segmentation threshold.

[0037] In the aforementioned technical solution, during a periodic patrol of the target lake, the drone completes its flight along a preset route, acquiring high-resolution visible light images covering the entire lake. These images contain three bands: blue, green, and red. Preprocessing steps convert the image data into surface reflectance data. For a typical pixel in the image, after correction, the reflectance in the green band (center wavelength approximately 555nm) is... The reflectance is 0.18 in the near-infrared band (center wavelength approximately 865 nm, acquired synchronously by a multispectral camera). The resolution was set to 0.02; then, using the UAV's POS (Positioning and Orientation System) data and ground control points, the image was geometrically corrected to give it accurate geographic coordinates, and the ground resolution (GSD) was determined to be 0.3 meters per pixel. Then, in the preprocessed image, the NDWI value is calculated pixel by pixel, using the following formula: For the example pixel above, its NDWI value is calculated to be 0.80; a higher... A value greater than 0 indicates that the pixel is likely a body of water, because water reflects light strongly in the green band and absorbs it strongly in the near-infrared band. To accurately separate water bodies from non-water bodies (such as vegetation, bare soil, and buildings) in a whole image, it is necessary to find an optimal segmentation threshold. The maximum entropy method is used to automatically determine this threshold, avoiding the arbitrariness and unsuitability of manually setting the threshold. This method believes that when the segmentation threshold divides the image into two categories, water bodies and non-water bodies, the optimal threshold is the one that maximizes the sum of the grayscale information entropy (NDWI value) of the pixels in these two categories. The larger the entropy value, the more uniform the distribution within the category, and the better the classification effect. All pixels in the NDWI image with an NDWI value greater than 0.45 are marked as water bodies, and the rest are marked as non-water bodies. This step generates a binary map, in which the boundary between water bodies and non-water bodies is the extracted instantaneous water boundary line. The system automatically stores the instantaneous area obtained from this monitoring into the database and compares it with the historical time series data of the lake.

[0038] As one embodiment of the present invention, the specific method for determining the instantaneous water area based on the normalized water index is as follows: Determining the adaptive segmentation threshold using the maximum entropy method When the NDWI value of any pixel in a visible light image is greater than the segmentation threshold When this happens, the pixel is identified as a water area pixel; the total number of all water area pixels is counted. Combined with the ground resolution of visible light imagery Calculate the instantaneous water area Calculation formula:

[0039] in, The unit is square meters.

[0040] In the above technical solution, the instantaneous water area of ​​the monitored lake is accurately calculated from a single visible light image acquired by a UAV using the Normalized Difference Water Index (NDWI) and the maximum entropy method. The UAV flies over the target lake at noon (when lighting conditions are good and there are no clouds), and the visible light camera (equipped with the ability to acquire green light and near-infrared bands, or working in conjunction with a multispectral camera) acquires an image covering the entire lake. The original resolution of this image is 5472×3648 pixels. By using GPS / IMU data recorded by the UAV and a small number of ground control points, and by performing orthorectification on the image to eliminate geometric distortion, the ground resolution (GSD) of the corrected image is accurately determined to the corresponding pixels.

[0041] By linearly mapping the NDWI values ​​of all pixels in the entire image (theoretically ranging from -1 to 1) to gray levels of 0-255; NDWI=-1 corresponds to gray level 0, NDWI=1 corresponds to gray level 255, the frequency of pixels appearing at each gray level is counted to form a histogram; a loop is set to traverse from gray level 1 to 254, and the NDWI value corresponding to the current gray level is used as a candidate segmentation threshold each time; for each candidate threshold corresponding to the NDWI value, the total entropy reaches the global maximum value, and the adaptive segmentation threshold of the image is determined. Based on the determined segmentation threshold, the entire image is judged. If the NDWI value of a pixel is greater than the segmentation threshold, the pixel is marked as 1; otherwise, it is marked as 0. The total number of water area pixels is counted, and the instantaneous water area is calculated based on the total number of water area pixels to obtain the current instantaneous water area of ​​the lake.

[0042] Please see Figure 2 As shown, in one embodiment of the present invention, the method for constructing the water level-area mathematical relationship model of the monitored water area in step S2 includes: S21. Collect water level observation data of the monitored water area at multiple historical periods. and the corresponding measured water area This constitutes a water level-area sample dataset. ; S22. Use nonlinear regression to fit the sample dataset and establish the water level-area function relationship:

[0043] in, For real-time hydrological water level data, For the corresponding water area, , , Parameters for model fitting; S23. Based on the water level-area function relationship, the actual water surface area... Uniform calibration to the preset reference water level Standardized area The calibration formula is:

[0044] in, This represents the real-time water level corresponding to the moment of aerial photography. This refers to the area corresponding to the current water level, calculated based on the water level-area function. This represents the area corresponding to the benchmark water level, calculated based on the water level-area function.

[0045] In the above technical solution, historical data is collected, a nonlinear regression model is constructed, and the instantaneous water surface area obtained at different water levels is uniformly calibrated to a unified reference water level, thereby eliminating the impact of water level fluctuations on area monitoring and achieving comparability of data from different periods. Firstly, taking a medium-sized reservoir in a hilly area in southern my country as an example, this reservoir has an irregular shape and significant variations in bank slope. To establish its water level-area relationship model, historical data from the past five years was collected. Water level data was obtained daily at 8:00 AM from the reservoir management office's automatic hydrological monitoring system, in meters (Yellow Sea Height System). Corresponding to the above water level observation day, the reservoir's water surface area, in square kilometers (km²), was obtained using high-resolution satellite remote sensing images (such as Sentinel-2, GF-2) or historical archived images from UAVs, through manual interpretation or semi-automatic extraction. 2 Then, after removing time points due to poor image quality or obvious data anomalies, a total of 60 valid water level-area sample pairs were obtained; thus forming the water level-area sample dataset. ,in ; Then, considering that the reservoir basin shape of natural water bodies is usually irregularly funnel-shaped, the relationship between water level and area is not a simple linear relationship, but rather closer to a quadratic curve relationship; therefore, a quadratic polynomial is chosen as the regression model: Using Python's scientific computing libraries or professional statistical analysis software, the above 60 sample datasets can be processed. The input model is fitted using the least squares method for nonlinear regression. After iterative calculation, the program successfully converges and obtains the optimal solution for the model parameters. Therefore, the established water level-area function relationship is used to explain the area change and the fitting effect is used for subsequent area calibration. Finally, in a new drone patrol mission, the drone captured visible light images of the reservoir and extracted the instantaneous water surface area using the method in step S1. Simultaneously, it acquired real-time water level data at the time of the aerial photography from the hydrological telemetry system. To facilitate comparisons of annual and multi-year area changes, a unified benchmark water level was established. In this example, the reservoir's normal storage level was selected as the benchmark water level. Through calculation formula Calculate the standardized area Real-time water level Substitute into the functional relationship Obtained through calculation.

[0046] Please see Figure 2 As shown, in one embodiment of the present invention, the method for obtaining the water surface area change rate and change patch data of the monitored water area over many years and multiple periods in step S3 includes: S31. Based on the standardized area obtained from multiple consecutive monitoring periods, calculate the... The monitoring cycle is relative to the first Water area change rate over monitoring periods The calculation formula is:

[0047] in, For the first Standardized area for each monitoring cycle for Standardized area for each monitoring cycle; S32. Perform spatial overlay analysis on the standardized areas of adjacent monitoring periods to mark the spatial areas where the water boundary changes, and generate change patch data. The change patch data includes the latitude and longitude information, location information, area information and change nature information of the change patches. S33. The sequence of water area change rates obtained over multiple years and periods. The change patch data serves as the basic input data for the time series analysis model.

[0048] In the above technical solution, the process of calculating the water surface area change rate and acquiring change patch data in step S3 is explained by calculating the area change rate based on standardized area and extracting change patches through spatial overlay analysis, providing basic input for the time series prediction model; and after processing in step S2, the standardized area of ​​the benchmark water level of the lake under multiple consecutive monitoring periods (set as one period every two weeks) is determined. Concurrently, each period also preserves a standardized water boundary vector map layer that has been geographically corrected and vectorized. Specifically: First of all, the The monitoring cycle is relative to the first Water area change rate over monitoring periods The calculation formula is Substitute the current period data into the time series data to calculate the rate of change for each period, thus forming a time series of rate of change: And this sequence is used for subsequent LSTM time series modeling; then step S32 performs spatial overlay analysis, by comparing two adjacent periods ( and The standardized water area vector map layers are used for spatial overlay analysis. Specific operations include acquiring water expansion areas (new water bodies) and water recession areas (reduced water bodies), generating change vector map layers for each type of area; after spatial overlay analysis, attribute values ​​are assigned to the change patches, automatically assigning attribute fields and data information to each change patch. For example, after overlaying in phase 10, an expansion-type change patch (area: 12,450 m²) is extracted in the northeast corner of the lake. 2 1. Nature of change: Expansion (possibly due to rainfall and water storage or partial inundation of tidal flats) Latitude and longitude: (117.2563, 31.8421). If the retreating pattern appears on the southwest side of the lake, it suggests that it may be due to shoreline erosion or human reclamation. Finally, in step S33, the time series of water area change rate and the multi-period change patch dataset are jointly stored in the monitoring database as input to the time series analysis model in step S3. The water area change rate time series... The input sequence for LSTM model construction (time window sliding sampling) includes a multi-period change map dataset containing a set of change map vectors, spatial location (latitude and longitude), area and water change properties (expansion or retreat) between adjacent periods. This combined input enables the time series model to not only predict the overall area change trend, but also to assist in the spatial identification of risk areas by combining historical change spatial distribution. This process provides a basis for dynamic map generation in step S4.

[0049] In one embodiment of the present invention, step S3 uses a time series analysis model to predict the future area change trend of the water body, and the method for obtaining the prediction coefficients includes: S34. Construct an LSTM (Long Short-Term Memory) network model. The LSTM model includes an input layer, an LSTM hidden layer, and an output layer. S35. Analyze the historical time series of water area change rates. The training sample set is constructed by sliding a time window, using the input sequence as input; the input of each training sample is continuous. Change rate data at each historical time step Output label is future Change rate data at each time step ; S36. The LSTM model is trained using the backpropagation algorithm and the adaptive moment estimator optimizer, with the goal of minimizing the mean squared error loss function between the predicted value and the true value, to obtain the trained LSTM prediction model. S37, Move the most recent consecutive The water area change rate data at each time step is input into the trained LSTM prediction model, and the model outputs a predicted area change rate sequence for the next preset time step, which serves as the prediction coefficient.

[0050] In the above technical solution, an LSTM model is constructed and trained to predict the future area change rate. After calculation in step S31, a standardized area change rate sequence for the lake over 104 consecutive weeks has been obtained. To improve the training stability and convergence speed of the LSTM model, the rate of change data is normalized using Min-Max, mapping the data to the [0,1] interval. First, step S34 uses a deep learning framework to construct an LSTM model and sets hyperparameters, including time window length, prediction step size, batch size, number of training epochs, and early stopping. Training stops when the validation set loss no longer decreases after 10 consecutive epochs. Then, in step S35, a model with a length of... A sliding window of 10 is used to extract training samples from the normalized rate of change sequence for each time step. Construct a training sample and input it. Normalized rate of change over 10 consecutive historical time steps Output This represents the normalized rate of change over the next four time steps. , and with Taking 14 as an example, its training samples are the input... This represents the normalized rate of change from week 4 to week 13; This represents the normalized rate of change from week 14 to week 17, and generates 94 samples from a total of 104 weeks of data processing. =11 to =104), divided according to the following proportions, with the training set consisting of 66 samples, and the time is... =11 to =76, the validation set was divided into 14 samples, and the time was... =77 to =90, the test set was divided into 14 samples, and the time was... =11 to =104.

[0051] Next, in step S36, the loss function in model training is determined by the mean squared error (MSE), and the training data is input into the LSTM model for batch iterative training. After each training round, the MSE loss on the validation set is calculated to monitor for overfitting. After 120 training rounds, the model converged, with a final MSE of 0.0032 on the training set, 0.0048 on the validation set, and 0.0051 on the test set, indicating good generalization ability. Finally, in step S37, by inputting the current time (end of week 104), the area change rate for the next four weeks (weeks 105-108) is predicted, using the normalized change rate of the most recent 10 consecutive weeks (weeks 95-104) as input. ,Will The trained LSTM model is input, and the model outputs normalized predictions for the next four time steps. The predicted values ​​are then converted back to the actual percentage change rate using previously saved normalization parameters. The LSTM model's final output includes the predicted area change rate sequence and the predicted absolute area value. The LSTM model successfully captures the long-term dependence and periodicity of lake area changes, and the prediction results show the lake area trend over future periods, which is consistent with the actual water temperature trend. The prediction coefficients will serve as key inputs for the generation of dynamic patches for risk area identification in step S4, providing data support for early warning.

[0052] In one embodiment of the present invention, the prediction coefficient includes the calculation formula for the predicted area change rate:

[0053] in, For the future Predicted rate of change of water area over a time step For time window The historical area change rate input sequence within, This is the set of training weight parameters for the LSTM model. Predictive coefficients also include risk level probability vectors The calculation formula is as follows:

[0054] in, For LSTM models in the future The hidden state vector output at each time step. and Here are the weight matrix and biases for the fully connected layer; , , , , These correspond to multiple preset risk levels of probability values, where 0 < < < < < <1, ,and ∈ ; The prediction coefficient also includes the absolute value of the predicted area. The calculation formula is as follows:

[0055] in, This refers to the standardized area for the current assessment period.

[0056] In the above technical solution, the current standardized area is in the 104th monitoring cycle. Historical input sequence: The normalized area change rate from week 95 to 104 is used to form the input vector. The trained LSTM weight set W (including forget gate, input gate, and output gate parameters) is used. Risk levels are defined as five levels: extremely low risk, low risk, medium risk, relatively high risk, and high risk. Predicting the future... Step area change rate The regression branch, the LSTM regression output, will Input a trained LSTM model to obtain the future nth Predicted normalized rate of change at each time step (taking k=1, i.e., week 105 as an example) Then, the actual rate of change is calculated through normalization. For example, the model output shows a predicted area change rate of 0.12%, indicating that under the standardization condition of the baseline water level, the lake surface area is expected to increase by approximately 12 ten-thousandths in the next monitoring period compared to the current period; LSTM at time... The output hidden state vector is denoted as , The value is 64, which corresponds to the number of units in the LSTM. This hidden state contains high-dimensional abstract features of historical temporal change patterns. A fully connected layer is connected in parallel on top of the LSTM. The calculation formula is... ,in This is the probability vector of risk level. This represents the weight matrix of the fully connected layer. ∈ bias ∈ , The risk level is determined based on the calculated risk level probability vector, and the risk level is determined by using the probability values ​​of multiple preset risk levels as a threshold: , , , , The risks corresponding to these values ​​include: extremely low risk level probability value, low risk level probability value, medium risk level probability value, relatively high risk level probability value, and high risk level probability value; furthermore, the method for identifying risk areas based on prediction coefficients in step S4 includes: Preset risk level thresholds, including high-risk probability thresholds. Medium risk probability threshold and low-risk probability threshold ; When the high-risk level probability in the risk level probability vector Greater than or equal to the high-risk probability threshold When the probability of a high-risk area is high, the area is classified as a high-risk area; when the probability of a high-risk level is high... Less than the high-risk probability threshold And the probability of medium risk level Greater than or equal to the medium risk probability threshold When the risk level probability vector is lower than the corresponding threshold, the area is classified as a medium-risk area; when all risk level probabilities in the risk level probability vector are lower than the corresponding threshold, the area is classified as a low-risk area; the scoring level corresponds to the risk level, with high-risk areas corresponding to the first scoring level, medium-risk areas corresponding to the second scoring level, and low-risk areas corresponding to the third scoring level; if the judgment is greater than the preset high-risk level threshold, the area is defined as a high-risk area.

[0057] In addition, based on the predicted rate of change and the current standardized area, the absolute value of the predicted area at the k-th step is calculated. Substitute the values ​​into the calculation to determine the result.

[0058] As one embodiment of the present invention, after generating the dynamic patches and rating levels of the risk area in step S4, a warning information generation step is also included: The dynamic patches are compared with pre-stored historical change patches and predicted dynamic patches; the predicted dynamic patches are the expected change areas within the future prediction period generated based on prediction coefficients. Calculate the deviation rate between the area of ​​the dynamic patch and the predicted area of ​​the dynamic patch. The calculation formula is:

[0059] in, This represents the actual area of ​​the dynamic patch. To predict the area of ​​dynamic patches; Preset deviation rate threshold When the deviation rate The absolute value is greater than or equal to the deviation rate threshold. When this occurs, the early warning mechanism is triggered, and early warning information is generated; Early warning information includes the warning level, warning type, location of the risk area, predicted evolution trend, and recommended response measures; the warning level is based on the deviation rate. The magnitude of the deviation rate is divided into several levels, and the larger the absolute value of the deviation rate, the higher the warning level.

[0060] In the above technical solution, by generating early warning information based on deviation rate, it is ensured that after generating dynamic patches and rating levels for risk areas, a graded early warning is triggered and complete early warning information is generated by comparing the area deviation between the actual dynamic patches and the predicted dynamic patches. The specific steps are as follows: First, based on the high-risk areas identified in step S4, dynamic patches for these areas are generated in the 105th monitoring period. Based on the prediction coefficients (predicted area change rate) output by the LSTM model... =+0.12%), combined with the spatial distribution patterns of historical change patches, the system generated a predicted dynamic patch for the 105th week using a spatial extrapolation algorithm; this patch represents the spatial area where water boundary changes are expected under the predicted trend; the area of ​​this predicted dynamic patch was automatically measured by GIS software. 8200 m 2 Then, after the actual UAV patrol mission in week 105 was completed, the system performed a spatial overlay analysis on the newly acquired standardized water boundary and the boundary from week 104, extracting the areas within the high-risk region that had actually changed, and generating actual dynamic patches; the area of ​​these actual dynamic patches was automatically measured by GIS software and found to be 10350 m². 2 And calculate the deviation rate between the actual dynamic patch area and the predicted dynamic patch area according to the preset deviation rate calculation formula: Substitute the above values ​​into the calculation. The calculation results show that the actual changed area exceeded the predicted changed area by 26.22%, a positive deviation, indicating that the area has expanded beyond expectations. Finally, threshold judgment and early warning triggering are performed. The system pre-sets three levels of deviation rate thresholds to determine the severity of the early warning trigger. Once the corresponding early warning threshold is exceeded, an early warning signal is triggered under that condition. For example, the first-level threshold is set at 10%. If this threshold is exceeded, a deviation log is recorded, but no early warning is triggered. The second-level threshold is 20%. If this second-level threshold is exceeded, a regular early warning is triggered. If it continues to exceed this threshold, a third-level threshold of 35% is reached. Once this third-level threshold is exceeded, an emergency early warning is triggered. Therefore, the above calculations... This indicates that if the warning occurs at the level 2 or level 3 threshold time, the system will determine that the warning mechanism has been triggered and classify the warning level as level 2. Therefore, the system will automatically generate structured warning information and push it to the management platform and relevant personnel terminals through the data interface.

[0061] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, and therefore described more simply; relevant parts can be referred to the descriptions of the method embodiments.

[0062] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in this application, they should all fall within the protection scope of the present invention.

Claims

1. A method for monitoring lake surface fluctuations based on unmanned aerial vehicles (UAVs), characterized in that, The method includes: S1. Conduct periodic patrols of the monitored water area according to the preset route to acquire multispectral and visible light images at multiple time intervals and extract the actual water surface area. S2. Based on the real-time hydrological and water level data of the monitored water area, establish a mathematical relationship model between water level and area of ​​the monitored water area, and uniformly calibrate the actual water surface area to a standardized area under a preset reference water level. S3. After obtaining the water surface area change rate and change patch data of the monitored water area over many years and multiple periods, the time series analysis model is used to predict the future area change trend of the water area and obtain the prediction coefficient. S4. Identify risk areas based on prediction coefficients, monitor risk areas, and generate dynamic maps and rating levels for risk areas; S5, and displays risk information in three dimensions based on dynamic patterns.

2. The method for monitoring lake surface fluctuations based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The method for extracting the water surface area is as follows: Based on the visible light image, the normalized water index is calculated, the instantaneous water boundary line of the monitored lake is extracted, the corresponding instantaneous water area is calculated, and it is compared with historical time series data to obtain the lake surface area fluctuation characteristics.

3. The method for monitoring lake surface fluctuations based on unmanned aerial vehicles (UAVs) according to claim 2, characterized in that, The normalized water index is calculated as follows: in, The reflectivity in the green light band, This refers to the reflectivity in the near-infrared band.

4. The method for monitoring lake surface fluctuations based on unmanned aerial vehicles (UAVs) according to claim 3, characterized in that, The specific method for determining the instantaneous water area based on the normalized water index is as follows: Determining the adaptive segmentation threshold using the maximum entropy method When the NDWI value of any pixel in the visible light image is greater than the segmentation threshold When this happens, the pixel is identified as a water area pixel; the total number of all water area pixels is counted. Combined with the ground resolution of the visible light image Calculate the instantaneous water area Calculation formula: in, The unit is square meters.

5. The method for monitoring lake surface fluctuations based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The construction methods for the water level-area mathematical relationship model of the monitored water area mentioned in step S2 include: S21. Collect water level observation data of the monitored water area at multiple historical periods. and the corresponding measured water area This constitutes a water level-area sample dataset. ; S22. The sample dataset is fitted using a nonlinear regression method to establish the water level-area function relationship: in, For real-time hydrological water level data, For the corresponding water area, , , Parameters for model fitting; S23. Based on the water level-area function relationship, the actual water surface area is... Uniform calibration to the preset reference water level Standardized area The calibration formula is: in, This represents the real-time water level corresponding to the moment of aerial photography. The area corresponding to the current water level is calculated based on the water level-area function relationship. The area corresponding to the reference water level is calculated based on the water level-area function relationship.

6. The method for monitoring lake surface fluctuations based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The methods for obtaining the water surface area change rate and change patch data of the monitored water area over many years and multiple periods, as described in step S3, include: S31. Based on the standardized area obtained from multiple consecutive monitoring periods, calculate the... The monitoring cycle is relative to the first Water area change rate over monitoring periods The calculation formula is: in, For the first Standardized area for each monitoring cycle for Standardized area for each monitoring cycle; S32. Perform spatial overlay analysis on the standardized areas of adjacent monitoring periods to mark the spatial areas where the water boundary changes, and generate change patch data. The change patch data includes the latitude and longitude information, location information, area information and change nature information of the change patches. S33. The sequence of water area change rates obtained over multiple years and periods. The change patch data serves as the basic input data for the time series analysis model.

7. The method for monitoring lake surface fluctuations based on unmanned aerial vehicles according to claim 6, characterized in that, Step S3 involves using a time series analysis model to predict future trends in water area changes. The methods for obtaining the prediction coefficients include: S34. Construct an LSTM (Long Short-Term Memory) network model, wherein the LSTM model includes an input layer, an LSTM hidden layer, and an output layer; S35. Analyze the historical time series of water area change rates. The training sample set is constructed by sliding a time window, using the input sequence as input; the input of each training sample is continuous. Change rate data at each historical time step Output label is future Change rate data at each time step ; S36. The LSTM model is trained using the backpropagation algorithm and the adaptive moment estimation optimizer, with the goal of minimizing the mean squared error loss function between the predicted value and the true value, to obtain the trained LSTM prediction model. S37, Move the most recent consecutive The water area change rate data at each time step is input into the trained LSTM prediction model, and the model outputs a predicted area change rate sequence for a future preset time step, which serves as the prediction coefficient.

8. The method for monitoring lake surface fluctuations based on unmanned aerial vehicles according to claim 7, characterized in that, The prediction coefficients include the formula for calculating the predicted area change rate: in, For the future Predicted rate of change of water area over a time step For time window The historical area change rate input sequence within, This is the set of training weight parameters for the LSTM model. The prediction coefficients also include a risk level probability vector. The calculation formula is as follows: in, For the LSTM model in the future The hidden state vector output at each time step. and Here are the weight matrix and biases for the fully connected layer; , , , , These correspond to multiple preset risk levels of probability values, where 0 < < < < < <1, ,and ∈ ; The prediction coefficient also includes the absolute value of the predicted area. The calculation formula is as follows: in, This refers to the standardized area for the current assessment period.

9. The method for monitoring lake surface fluctuations based on unmanned aerial vehicles according to claim 8, characterized in that, The method for identifying risk areas based on prediction coefficients as described in step S4 includes: Preset risk level thresholds, including high-risk probability thresholds. Medium risk probability threshold and low-risk probability threshold ; When the high-risk level probability in the risk level probability vector Greater than or equal to the high-risk probability threshold At that time, the area was identified as a high-risk area; When the probability of the high risk level Less than the high-risk probability threshold And the probability of medium risk level Greater than or equal to the aforementioned risk probability threshold At that time, the area was identified as a medium-risk area; When all risk level probabilities in the risk level probability vector are lower than the corresponding threshold, the area is determined to be a low-risk area. The rating levels correspond to the risk levels, with high-risk areas corresponding to the first rating level, medium-risk areas corresponding to the second rating level, and low-risk areas corresponding to the third rating level.

10. The method for monitoring lake surface fluctuations based on unmanned aerial vehicles according to claim 1, characterized in that, After generating dynamic patches and rating levels for risk areas as described in step S4, the process also includes a step of generating early warning information: The dynamic patch is compared with pre-stored historical change patches and predicted dynamic patches; the predicted dynamic patch is the expected change area within the future prediction period generated based on the prediction coefficient; Calculate the deviation rate between the area of ​​the dynamic patch and the area of ​​the predicted dynamic patch. The calculation formula is: in, The actual area of ​​the dynamic patch. The predicted area of ​​the predicted dynamic patch; Preset deviation rate threshold When the deviation rate The absolute value is greater than or equal to the deviation rate threshold. When this occurs, the early warning mechanism is triggered, and early warning information is generated; The early warning information includes the early warning level, early warning type, location of the risk area, predicted evolution trend, and recommended response measures; the early warning level is based on the deviation rate. The magnitude of the deviation rate is divided into several levels, and the larger the absolute value of the deviation rate, the higher the warning level.

Images