A method for constructing a water level prediction model based on image recognition

Abstract

The present invention discloses a method for constructing a water level prediction model based on image recognition, which relates to the technical field of water level monitoring, and includes the following steps: establishing a visual reference system for the region, constructing a multi-spectral image data fusion module, constructing a water surface edge detection module, designing an anomaly detection mechanism based on self-supervised learning, constructing a water level height estimation model based on image semantic segmentation, constructing a water level distribution prediction model combined with geographic information, constructing a data feedback model based on meteorological dynamics, and introducing a water level trend estimation module for short-term time series prediction. This method constructs a water level prediction model by introducing technologies such as multi-spectral image fusion, semantic segmentation, and dynamic height estimation, and realizes the accurate extraction and prediction of water level information under various environmental conditions. The robustness of the model under complex weather conditions is significantly improved, and at the same time, the deviation problems caused by different perspectives and optical distortions are solved through a dynamic adjustment factor and a geometric correction mechanism.

Description

Technical Field

[0001] The present invention relates to the technical field of water level monitoring, and specifically to a method for constructing a water level prediction model based on image recognition. Background Art

[0002] Currently, water level monitoring systems usually rely on Internet of Things sensors, radar, or ultrasonic ranging devices. However, these technologies face many challenges in practical applications: the hardware devices are vulnerable to the external environment (such as heavy rainfall, floating object blockage), and when deployed on a large scale, the maintenance cost is high. In recent years, water level monitoring technology based on image recognition has gradually emerged, which realizes water level monitoring and prediction by collecting images through cameras and combining algorithms. Compared with traditional devices, this method has the advantages of non-contact, high flexibility, and low cost.

[0003] In the prior art, in the case of insufficient light or strong reflection (such as at night, during heavy rain, or in haze), the resolution ability between the water surface and the markers significantly decreases for traditional single-spectral image methods. Many methods rely on traditional edge detection or simple segmentation algorithms and are difficult to cope with the interference brought by complex backgrounds (such as floating objects, reflective water surfaces). Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a method for constructing a water level prediction model based on image recognition to solve the problems raised in the above background art.

[0005] To achieve the above object, the present invention provides the following technical solution: A method for constructing a water level prediction model based on image recognition, including the following steps:

[0006] S1. In the target water area, establish a visual reference system for the area;

[0007] S2. In the visual reference system, construct a multi-spectral image data fusion module to fuse the image data collected from the target water area;

[0008] S3. In the visual reference system, construct a water surface edge detection module to locate the water surface edge of the target water area;

[0009] S4. Design an anomaly detection mechanism based on self-supervised learning to perform anomaly detection on the image data collected from the target water area;

[0010] S5. Construct a water level height estimation model based on image semantic segmentation to separate the water surface, markers, and other background areas of the target water area and estimate the water level height;

[0011] S6. Integrate the geographical information of the surrounding area of the target water area and construct a water level distribution prediction model combined with geographical information;

[0012] S7. Build a data feedback model based on meteorological dynamics, and access meteorological data for water level prediction calculations;

[0013] S8. Introduce a water level trend estimation module for short-term time series prediction to complete the construction of the water level prediction model.

[0014] To further optimize this technical solution, in step S1, by analyzing the terrain of the target water area and known water level markers, including bridge piers and water level posts, a visual reference system is designed;

[0015] The reference system determines the specific features of the water level markers in the image, including color changes and shape markings in height segments, and is established in three-dimensional space in combination with image data from multiple angles.

[0016] To further optimize this technical solution, in step S2, the multi-spectral image data fusion module is built with an adaptive weight allocation algorithm, which dynamically adjusts the credibility of different spectral data according to the acquisition environment. The spectral data includes visible light, infrared, and ultraviolet data, and the fused image can always reflect the real changes in the water level.

[0017] To further optimize this technical solution, the adaptive weight allocation algorithm dynamically adjusts the credibility of visible light, infrared, and ultraviolet data to ensure that the fused image can accurately reflect the real changes in the water level;

[0018] The adaptive weight allocation algorithm includes the following processes:

[0019] Input multi-spectral image data;

[0020] Calculate the signal-to-noise ratio of the image;

[0021] Calculate the environmental factor weights;

[0022] Normalize the weights;

[0023] Multi-spectral image fusion.

[0024] To further optimize this technical solution, in step S3, the water surface edge detection module is based on a deep reinforcement learning algorithm. The model learns how to detect the water surface edge in a complex background from a large amount of training data and dynamically adjusts the strategy during detection to improve the robustness of water surface detection and provide accurate basic data for water level prediction;

[0025] When the water surface ripples are large, the model gives priority to focusing on smooth areas;

[0026] When the light reflection is strong, the model chooses to avoid the highlighted areas.

[0027] To further optimize this technical solution, when the anomaly detection mechanism is in use, in step S4, it includes:

[0028] When the water level marker is covered by plants or the camera is blurred by raindrops, the anomaly detection mechanism learns the intrinsic distribution characteristics of the data from normal water level images, actively issues a warning when detecting abnormal input, and completes the missing image information through the generative adversarial network GAN.

[0029] To further optimize this technical solution, in step S5, the water level height estimation model includes:

[0030] Define the water level height :

[0031] ;

[0032] Among them,

[0033] : The vertical pixel coordinate of the water surface area, the pixel row number from the top of the image to the water surface junction area;

[0034] : The vertical pixel coordinate of the marker reference point, the pixel row number from the top of the image to the bottom of the marker;

[0035] : The mapping coefficient between the pixels of the image height and the real height;

[0036] : The dynamic environment adjustment factor, used to correct the errors caused by optical distortion or environmental interference;

[0037] Segmentation mapping:

[0038] The semantic segmentation network divides the image into three types of regions: water surface, marker, and background, and outputs the mapping;

[0039] Geometric correction:

[0040] Automatically adapts to images with different perspectives and camera positions to ensure the accuracy of unit conversion;

[0041] Dynamic environment adjustment factor:

[0042] In complex environments including strong light reflection and rain blurring, there are errors in segmentation, and the dynamic adjustment factor is used to correct the water level estimation.

[0043] To further optimize this technical solution, in the segmentation mapping, the following mapping is output:

[0044] : Pixel point The classification probability mapping of, where Represents the probability of the water surface, Represents the probability of the marker;

[0045] Water surface boundary line extraction:

[0046] ;

[0047] Among them, the threshold is used to determine whether a pixel point belongs to the water surface;

[0048] Landmark reference extraction:

[0049] ;

[0050] Through segmentation mapping, it is used to ensure that the model accurately extracts the boundary area between the water surface and the landmark from the complex background.

[0051] To further optimize this technical solution, in step S7, the data feedback model is used to introduce factors such as precipitation, wind speed, and temperature into the calculation of water level prediction;

[0052] The model inputs meteorological data in real time and affects the final water level prediction result in a weighted manner, and automatically optimizes the dependence on meteorological data according to environmental changes or seasonal precipitation patterns.

[0053] To further optimize this technical solution, in step S8, the water level trend estimation module takes the water level data output by image recognition as input, combines it with historical water level data, and predicts the future water level change trend;

[0054] The model combines the sliding window technology during training to extract the short-term dynamic characteristics of the water level, provides accurate trend prediction results, and thus predicts the water level trend within the next 1 - 3 hours.

[0055] Compared with the prior art, the present invention provides a method for constructing a water level prediction model based on image recognition, which has the following beneficial effects:

[0056] This method for constructing a water level prediction model based on image recognition constructs a water level prediction model by introducing technologies such as multi-spectral image fusion, semantic segmentation, and dynamic height estimation, and realizes the accurate extraction and prediction of water level information under various environmental conditions. Compared with the prior art, the robustness of the model under complex weather conditions is significantly improved, and at the same time, the deviation problems caused by different perspectives and optical distortions are solved through the dynamic adjustment factor and geometric correction mechanism. The model has higher applicability, accuracy, and environmental adaptability, is applicable to multiple fields such as intelligent water resource management and flood warning, and effectively reduces the hardware maintenance cost and improves the monitoring efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is a schematic flow chart of a method for constructing a water level prediction model based on image recognition proposed by the present invention;

[0058] Figure 2 Schematic diagram of the adaptive weight allocation algorithm in a water level prediction model construction method based on image recognition proposed by the present invention;

[0059] Figure 3 Schematic diagram of the water level height estimation model in a water level prediction model construction method based on image recognition proposed by the present invention. Specific implementation manner

[0060] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0061] Embodiment 1:

[0062] Please refer to Figure 1 , a water level prediction model construction method based on image recognition, including the following steps:

[0063] S1. Establish a visual reference system for the area

[0064] In this embodiment, by analyzing the terrain of the target water area and known water level markers, including bridge piers and water level posts, a visual reference system is designed;

[0065] The reference system determines the specific features of the water level markers in the image, including color changes and shape markings in height segments, and is established in a three-dimensional space in combination with image data from multiple angles.

[0066] Accurate water level reference markers are required for water level prediction, and the markers in each target area may be different. By establishing a visual reference system, it can be ensured that the model can identify the basis for water level changes in this area. For example, the red, yellow, and green segmented markings on the bridge pier can be marked as key features, and at the same time, possible visual interferences such as algae and dirt need to be eliminated. The process of constructing the reference system includes taking panoramic images and close-up images of the water area, and applying deep learning technology to perform feature annotation and classification on the water level markers. This system will serve as the reference framework for the entire model to ensure the regional specificity and accuracy of subsequent predictions.

[0067] S2. Construct a multi-spectral image data fusion module

[0068] Due to the fact that light conditions (such as at night, in fog, or during heavy rain) may cause the quality of visible light images to decline, multi-spectral data acquisition technology is introduced. Infrared images can detect thermal signals under low-light or no-light conditions, while ultraviolet light is sensitive to the reflection of organic matter or floating objects in water. By fusing this spectral data with visible light images, stable water level information can be obtained under various climate conditions.

[0069] In this embodiment, the multi-spectral image data fusion module is built-in with an adaptive weight allocation algorithm, which dynamically adjusts the credibility of different spectral data according to the acquisition environment. The spectral data includes visible light, infrared, and ultraviolet data, and the fused image can always reflect the true change of the water level.

[0070] The adaptive weight allocation algorithm dynamically adjusts the credibility of visible light, infrared, and ultraviolet data to ensure that the fused image can accurately reflect the true change of the water level.

[0071] S3. Construct a water surface edge detection module

[0072] In this embodiment, the positioning of the water surface edge is a key step in predicting the water level. However, due to interference from reflection, ripples, and floating objects, traditional edge detection algorithms (such as the Canny algorithm) may not be reliable enough. The water surface edge detection module is based on a deep reinforcement learning algorithm. The model learns how to detect the water surface edge in a complex background from a large amount of training data and dynamically adjusts the strategy during detection to improve the robustness of water surface detection and provide accurate basic data for water level prediction.

[0073] When the water surface ripples are large, the model gives priority to focusing on smooth areas.

[0074] If the standard deviation of the gray level or texture change in the water surface area of the image exceeds a preset threshold (for example, the standard deviation is greater than 50), it is considered that the water surface ripples are large. At the same time, in the frequency domain, if the amplitude of the ripples exceeds a set threshold (for example, the frequency distribution of the ripples is higher than a certain standard value, such as 5 Hz), it can also be considered large ripples.

[0075] When the light reflection is strong, the model chooses to avoid the high-brightness area.

[0076] In an RGB image, if the pixel brightness (such as the gray value or the maximum value of the R, G, B three channels) is greater than a preset threshold (such as 200), it can be considered that the area has strong reflection. If the brightness value of a certain area exceeds the brightness of 99% of the pixels in the image (such as exceeding the 95th percentile), it can also be considered that the light reflection is strong.

[0077] S4. Design an anomaly detection mechanism based on self-supervised learning

[0078] During actual operation, the model may be affected by problems such as equipment occlusion and damage, resulting in abnormal data input.

[0079] In this embodiment, when the anomaly detection mechanism is in use, it includes:

[0080] When the water level marker is covered by plants or the camera is blurred by raindrops, the anomaly detection mechanism learns the intrinsic distribution characteristics of the data from normal water level images, actively issues a warning when detecting abnormal input, and completes the missing image information through the generative adversarial network GAN.

[0081] S5. Construct a water level height estimation model based on image semantic segmentation

[0082] Calculating the water level height directly in the image requires accurate segmentation of the water surface area. Perform pixel-level segmentation on the input image to separate the water surface, marker, and other background areas. The segmentation result can be directly used for the quantitative calculation of the water level height. For example, calculate the actual water level through the height markings on the marker. The model needs to be trained on diverse data under different lighting and weather conditions to ensure the stability and universality of the segmentation result.

[0083] S6. Construct a water level distribution prediction model combining geographical information

[0084] In this embodiment, the water level change is not only affected by the local environment but also closely related to the topography and landforms of the surrounding areas. For example, precipitation in an upstream area may cause the water level in the downstream to rise. Therefore, by integrating GIS data, including river networks, terrain elevation differences, historical precipitation records, etc., establish a water level distribution prediction model to predict the water level based on the topographical and landform information.

[0085] S7. Construct a data feedback model based on meteorological dynamics

[0086] In this embodiment, the water level prediction not only depends on image data but is also closely related to meteorological factors. For example, heavy rain will significantly increase the water level, while continuous high temperature may cause the water level to drop. The data feedback model is used to introduce factors such as precipitation, wind speed, and temperature into the calculation of water level prediction;

[0087] The model inputs meteorological data in real time, affects the final water level prediction result in a weighted manner, and automatically optimizes the dependence on meteorological data according to environmental changes or seasonal precipitation patterns.

[0088] S8. Introduce a water level trend estimation module for short-term time series prediction

[0089] In this embodiment, the water level trend estimation module takes the water level data output by image recognition as input, combines it with historical water level data, and predicts the future water level change trend, thus completing the construction process of the water level prediction model.

[0090] The water level prediction model extracts short-term dynamic features of the water level by combining the sliding window technique during training, providing accurate trend prediction results, thereby predicting the water level trend within the next 1 - 3 hours.

[0091] Example 2:

[0092] Please refer to Figure 2 and Figure 3 Based on the method for constructing a water level prediction model based on image recognition described in Example 1, in step S2, design an adaptive weight allocation algorithm.

[0093] Step 1: Input multispectral image data

[0094] Set:

[0095] : Visible light image;

[0096] : Infrared light image;

[0097] : Ultraviolet light image.

[0098] Collect the environmental information as , including the following environmental features:

[0099] Light intensity ;

[0100] Weather conditions (such as sunny, rainy, foggy);

[0101] Time period (day or night).

[0102] Step 2: Calculate the signal-to-noise ratio of the image

[0103] For each spectral image, calculate the signal-to-noise ratio , to quantify the credibility of each image:

[0104]

[0105] Effective pixel intensity: Refers to the pixel intensity in the area of the water surface or water level marker in the image (achieved through preliminary segmentation).

[0106] Noise intensity: Refers to the standard deviation of the pixel intensity in the background area of the image.

[0107] Respectively obtain .

[0108] are respectively the signal-to-noise ratio of the visible light image, the signal-to-noise ratio of the infrared light image, and the signal-to-noise ratio of the ultraviolet light image.

[0109] Step 3: Calculate the environmental factor weights

[0110] Based on the environmental characteristics , determine the initial weights of the spectral types :

[0111] Visible light weight :

[0112]

[0113] When is higher (daytime, strong light environment), the importance of visible light is high, and the weight tends to 1.

[0114] When is lower (night, low light environment), the weight decreases.

[0115] Infrared light weight :

[0116]

[0117] In low light or rainy days, at night, the infrared weight increases.

[0118] and are gain functions based on weather and time.

[0119] Ultraviolet light weight :

[0120]

[0121] In foggy weather, the ultraviolet weight decreases; in sunny weather, the weight increases.

[0122] is a gain function based on the visibility (haze) in the environment.

[0123] Among them, is a tuning parameter.

[0124] Step 4: Normalize the weights

[0125] Combine the signal-to-noise ratio and the environmental weights.

[0126] Step 5: Multispectral image fusion

[0127] Based on the calculated weights, weight and fuse the three spectral images into a final image :

[0128]

[0129] The model adjusts weights in real time based on environmental features and signal-to-noise ratio, adapts to different light, weather, and time conditions, has strong adaptability to complex conditions such as low light, fog, and rain, combines signal-to-noise ratio and environmental factor weights, avoids direct reliance on complex deep learning calculations, and greatly improves real-time performance.

[0130] When the model is applied:

[0131] During sunny days: The weight is large, mainly relying on visible light images.

[0132] At night: The weight is large, and the infrared image contributes more.

[0133] During fog: Increase the weight of the ultraviolet image to capture the water surface reflection characteristics.

[0134] Design the water level height estimation model in step S5.

[0135] The water level height estimation model includes the following processes:

[0136] Define the water level height

[0137] ;

[0138] Among them,

[0139] : The vertical pixel coordinate of the water surface area, the pixel row number from the top of the image to the water surface junction area;

[0140] : The vertical pixel coordinate of the landmark reference point, the pixel row number from the top of the image to the bottom of the landmark;

[0141] : The mapping coefficient between the pixels of the image height and the real height;

[0142] : The dynamic environment adjustment factor, used to correct errors caused by optical distortion or environmental interference.

[0143] Segmentation mapping

[0144] The semantic segmentation network divides the image into three types of regions: water surface, landmark, and background, and outputs the following mapping:

[0145] : The classification probability mapping of pixel point Among them, represents the probability of the water surface, represents the probability of the landmark;

[0146] Water surface boundary line extraction:

[0147] ;

[0148] Among them, the threshold is used to determine whether a pixel point belongs to the water surface;

[0149] Marker reference extraction:

[0150] ;

[0151] Through segmentation mapping, it is used to ensure that the model accurately extracts the boundary area between the water surface and the marker from the complex background.

[0152] Geometric correction

[0153] is the conversion coefficient between the image pixel height and the real height, which can be dynamically calculated by the following formula:

[0154]

[0155] Among them,

[0156] : the real height of the marker (known).

[0157] : the vertical pixel coordinate of the top of the marker;

[0158] : the vertical pixel coordinate of the marker reference point.

[0159] Automatically adapts to images with different perspectives and camera positions to ensure the accuracy of unit conversion.

[0160] Dynamic environment adjustment factor

[0161] In complex environments including strong light reflection and rain blurring, there are errors in segmentation, and the dynamic adjustment factor is used to correct the water level estimation.

[0162] When this model is used, it includes:

[0163] Step 1: Semantic segmentation to extract the water surface and marker areas

[0164] Perform pixel-level classification on the input image to generate a classification probability map . Extract the water surface boundary points and the marker reference points through the threshold method.

[0165] Step 2: Determine the mapping between the pixel and real height of the marker

[0166] According to the known true height of the marker , calculate the conversion coefficient , which is used to deduce the true height from the pixel height.

[0167] Step 3: Calculate the base water level height

[0168] Divide the difference between the water surface pixel position and the marker reference pixel position by to obtain the base water level height.

[0169] Step 4: Dynamically adjust the water level height

[0170] Calculate the adjustment factor according to the image brightness (ambient light interference) and edge uncertainty , and correct the base water level height to obtain the final result.

[0171] This model utilizes geometric correction and dynamic adjustment factors, and is suitable for a variety of acquisition environments, including different perspectives, lighting conditions, and weather states. With the help of the semantic segmentation network, the boundary region between the water surface and the marker can be accurately extracted, avoiding the errors of traditional edge detection methods.

[0172] When this model is applied:

[0173] The true height of the marker is 2 meters;

[0174] The pixel positions of the top and bottom of the marker are 100 and 300 respectively;

[0175] The water surface pixel position is 299.95;

[0176] The dynamic adjustment factor is -0.05 (correction for the downward shift of the water surface caused by light reflection).

[0177] Calculate :

[0178]

[0179] Calculate the water level height:

[0180]

[0181] The final water level height is -5.05 meters (the water level is 5.05 meters downward relative to the marker reference point position).

[0182] The beneficial effects of the present invention are:

[0183] The method for constructing a water level prediction model based on image recognition constructs a water level prediction model by introducing technologies such as multispectral image fusion, semantic segmentation, and dynamic height estimation, and realizes the accurate extraction and prediction of water level information under various environmental conditions. Compared with the prior art, the robustness of the model under complex weather conditions is significantly improved, and at the same time, the deviation problems caused by different perspectives and optical distortions are solved through a dynamic adjustment factor and a geometric correction mechanism. The model has higher applicability, accuracy, and environmental adaptability, and is applicable to multiple fields such as intelligent water resource management and flood warning, effectively reducing the hardware maintenance cost and improving the monitoring efficiency.

[0184] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, without conflict, those skilled in the art can combine and combine the different embodiments or examples and different implementations described in this specification.

[0185] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a water level prediction model based on image recognition, characterized in that: The following steps are involved: S1. Establish a regional visual reference system in the target waters; S2. In the visual benchmark system, a multispectral image data fusion module is constructed to fuse the image data collected from the target water area; S3. In the visual reference system, a water surface edge detection module is constructed to locate the water surface edge of the target water area; S4. Design an anomaly detection mechanism based on self-supervised learning to perform image anomaly detection on the image data collected from the target water area; S5. Construct a water level estimation model based on image semantic segmentation to separate the water surface, landmarks and other background areas of the target water area to estimate the water level; The water level estimation model includes: Define water level : ； in, : The vertical pixel coordinate of the water surface area, the pixel row number from the top of the image to the water surface boundary area; : The vertical pixel coordinate of the marker reference point, the pixel row number from the top of the image to the bottom of the marker; : The conversion coefficient between image pixel height and real height is dynamically calculated by the following formula: ； in, : The real height of the landmark; : The vertical pixel coordinate of the top of the marker; : The vertical pixel coordinate of the reference point of the marker; : Dynamic environment adjustment factor, used to correct errors caused by optical distortion or environmental interference; Split Mapping: The semantic segmentation network divides the image into three regions: water surface, landmarks, and background, and outputs a map; Geometry Correction: Automatically adapt images of different viewing angles and camera positions to ensure the accuracy of unit conversion; Dynamic Environment Adjustment Factor: In complex environments including strong light reflection and rain blur, segmentation has errors, and the dynamic adjustment factor is used to correct the water level estimation; In the segmentation mapping, the following mapping is output: : Pixels The classification probability map of represents the probability of water surface, represents the probability of a marker; Water surface boundary extraction: ； Among them, the threshold Used to determine whether a pixel belongs to the water surface; Marker Benchmark Extraction: ； Segmentation mapping is used to ensure that the semantic segmentation network model can accurately extract the boundary area between the water surface and the landmark from the complex background; S6. Integrate the geographic information of the surrounding areas of the target waters and construct a water level distribution prediction model combining the geographic information; S7. Construct a data feedback model based on meteorological dynamics and access meteorological data to calculate water level prediction; S8. Introduce the water level trend estimation module of short-term time series prediction to complete the construction of the water level prediction model.

2. The method for constructing a water level prediction model based on image recognition according to claim 1, characterized in that: In the step S1, a visual reference system is designed by analyzing the topography of the target water area and known water level markers, including bridge piers and water level columns; The benchmark system determines the specific features of water level markers in the image, including highly segmented color changes, shape markings, and is established in three-dimensional space by combining multi-angle image data.

3. The method for constructing a water level prediction model based on image recognition according to claim 1, characterized in that: In step S2, the multispectral image data fusion module has a built-in adaptive weight distribution algorithm, which dynamically adjusts the credibility of different spectral data according to the acquisition environment. The spectral data includes visible light, infrared and ultraviolet data. The fused image can always reflect the real changes in the water level.

4. The method for constructing a water level prediction model based on image recognition according to claim 3 is characterized in that: The adaptive weight distribution algorithm dynamically adjusts the credibility of visible light, infrared and ultraviolet data to ensure that the fused image can accurately reflect the real changes in water level; The adaptive weight allocation algorithm includes the following processes: Input multispectral image data; Calculate the image signal-to-noise ratio; Calculate the weights of environmental factors; Normalized weights; Multispectral image fusion.

5. The method for constructing a water level prediction model based on image recognition according to claim 1, characterized in that: In step S3, the water surface edge detection module is based on a deep reinforcement learning algorithm. The water surface edge detection module learns how to detect the water surface edge in a complex background from a large amount of training data, and dynamically adjusts the strategy during detection to improve the robustness of water surface detection and provide accurate basic data for water level prediction; When the water surface ripples are large, the standard deviation of the grayscale or texture change of the water surface area in the image exceeds the preset threshold, then the water surface ripples are considered to be large; or in the frequency domain, the amplitude of the ripples exceeds the set threshold, then the water surface ripples are considered to be large; the water surface edge detection module focuses on the smooth area; When the light reflection is strong, in the RGB image, if the pixel brightness is greater than a preset threshold, it is considered that the light reflection is strong; the water surface edge detection module chooses to avoid the highlight area.

6. The method for constructing a water level prediction model based on image recognition according to claim 1, characterized in that: In step S4, when the anomaly detection mechanism is used, it includes: When the water level markers are covered by plants or the camera is blurred by raindrops, the anomaly detection mechanism learns the intrinsic distribution characteristics of the data from normal water level images, actively issues warnings when abnormal inputs are detected, and completes the missing image information through the generative adversarial network (GAN).

7. The method for constructing a water level prediction model based on image recognition according to claim 1, characterized in that: In step S7, the data feedback model is used to introduce factors such as precipitation, wind speed, and temperature into the calculation of water level prediction; The data feedback model accesses meteorological data in real time, affects the final water level prediction results in a weighted manner, and automatically optimizes the degree of dependence on meteorological data based on environmental changes or seasonal precipitation patterns.

8. The method for constructing a water level prediction model based on image recognition according to claim 1, characterized in that: In step S8, the water level trend estimation module uses the water level data output by the image recognition as input, combines it with the historical water level data, and predicts the future water level change trend; The water level trend estimation module combines the sliding window technology in training to extract the short-term dynamic characteristics of the water level and provide accurate trend prediction results, thereby predicting the water level trend in the next 1-3 hours.

Images