Intelligent monitoring method and system for cyanobacteria bloom outbreak

Abstract

This application relates to the field of data processing technology and discloses an intelligent monitoring method and system for cyanobacterial blooms. The method includes: collecting surface spectral and underwater particle size data and performing atmospheric correction; calculating the normalized algal index and blue-green band ratio, then inputting the calculated data into a U-Net network to obtain the bloom coverage area; integrating and interpolating the particle size data to obtain the vertical profile distribution curve and calculating the surface enrichment and buoyancy trend index; establishing the surface-underwater correlation through random forest regression training; and inputting the multidimensional features into a CNN-LSTM model to predict the bloom probability and determine the warning level. This application solves the problems of insufficient effective fusion of surface spectral data and underwater vertical profile data, which prevents a comprehensive depiction of the three-dimensional structure of cyanobacterial blooms, and the lack of a multi-source data time-series analysis model, which leads to insufficient timeliness of bloom warnings. It improves the spatial integrity and early warning lead time of cyanobacterial bloom monitoring.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to an intelligent monitoring method and system for cyanobacterial blooms. Background Technology

[0002] Eutrophication is severe in some lakes in my country, with frequent cyanobacterial blooms posing a serious threat to water source safety and the health of aquatic ecosystems. Traditional cyanobacterial monitoring methods based on manual sampling and laboratory analysis are limited in spatiotemporal coverage and require significant manpower and resources. While satellite remote sensing can rapidly monitor large-scale cyanobacterial blooms, it is difficult to obtain continuous and effective monitoring data due to cloud cover and low temporal resolution. Ground-based remote sensing technology can operate continuously under all-weather conditions. Through multispectral cameras installed at fixed monitoring stations, it enables automatic monitoring of surface cyanobacteria across multiple time phases, directions, and scenes. Deep learning image recognition algorithms are used to process multispectral images in real time, automatically extracting the distribution areas of surface cyanobacteria. Simultaneously, the LISST laser scattering in-situ sensor can measure vertical profile indicators such as algal biomass concentration, particle size distribution, and cell density at different water depths in real time, monitoring the vertical distribution and migration dynamics of underwater algae.

[0003] However, current ground-based remote sensing technologies can only acquire two-dimensional distribution information of the water surface, while underwater in-situ sensor data lacks effective correlation analysis with water surface spectral data, resulting in fragmented monitoring information. Ground-based remote sensing can only reflect the spectral characteristics of the water surface and cannot capture the vertical distribution and migration dynamics of underwater algae, while underwater sensors can only provide vertical profile data of a single point, making it difficult to reflect the spatial distribution pattern of the entire water area. This separation of surface and underwater data makes it difficult for existing monitoring methods to comprehensively characterize the three-dimensional structure of cyanobacterial blooms and effectively identify the dynamic process of cyanobacteria migrating and accumulating from underwater to the water surface, thus leading to insufficient timeliness in early warning of cyanobacterial blooms.

[0004] Due to the lack of correlation modeling between surface spectral characteristics and the vertical distribution of underwater biomass, existing technologies cannot infer the distribution of algae at different depths using surface remote sensing data, nor can they utilize underwater vertical profile information to correct and enhance the accuracy of surface monitoring. Furthermore, due to the lack of a time-series analysis model integrating multi-dimensional surface and underwater features, existing technologies struggle to capture the temporal evolution of cyanobacterial concentrations and key dynamic characteristics such as surface enrichment and upward migration, making it impossible to establish a cyanobacterial bloom prediction mechanism based on multi-source data fusion. More critically, the lack of an early warning and decision-making system that intelligently integrates surface spectral indices, underwater biomass vertical distribution characteristics, and temporal change patterns prevents existing technologies from identifying early signs of cyanobacterial blooms, hindering sufficient time for emergency response and directly impacting the effectiveness of cyanobacterial bloom control and management. Summary of the Invention

[0005] This application provides an intelligent monitoring method and system for cyanobacterial blooms, which solves the problem that the lack of effective fusion between surface spectral data and underwater vertical profile data makes it impossible to fully characterize the three-dimensional structure of cyanobacterial blooms, and the problem that the lack of a multi-source data time series analysis model leads to insufficient early warning timeliness of blooms, thereby improving the spatial integrity and early warning lead time of cyanobacterial bloom monitoring.

[0006] Firstly, this application provides an intelligent monitoring method for cyanobacterial blooms, the intelligent monitoring method for cyanobacterial blooms comprising:

[0007] Step S1: Collect water surface spectral reflectance data and underwater particle size volume concentration distribution data, and perform atmospheric correction processing on the water surface spectral reflectance data to obtain surface reflectance data;

[0008] Step S2: Calculate the normalized algal index and blue-green band ratio based on the surface reflectance data. Input the normalized algal index, the blue-green band ratio, and the surface reflectance data into a U-Net convolutional neural network for segmentation processing to obtain the algal bloom coverage area.

[0009] Step S3: Integrate the particle size volume concentration distribution data to obtain the biological volume concentration at each depth layer, interpolate the biological volume concentration at each depth layer to obtain the vertical profile distribution curve, and calculate the surface enrichment and floating trend index of the vertical profile distribution curve.

[0010] Step S4: Extract the normalized algal index at the location corresponding to the algal bloom coverage area and the biological volume concentration at each depth layer to form a registration sample pair, and train the registration sample pair using random forest regression to obtain the surface-underwater correlation model.

[0011] Step S5: Input the normalized algal index, the surface enrichment degree, and the rising trend index into the CNN-LSTM model to obtain the algal bloom probability, and determine the warning level based on the algal bloom probability and the rising trend index.

[0012] Secondly, this application provides an intelligent monitoring system for cyanobacterial blooms, the intelligent monitoring system for cyanobacterial blooms comprising:

[0013] The correction module is used to collect water surface spectral reflectance data and underwater particle size volume concentration distribution data, and to perform atmospheric correction processing on the water surface spectral reflectance data to obtain surface reflectance data.

[0014] The segmentation module is used to calculate the normalized algae index and the blue-green band ratio based on the surface reflectance data, and input the normalized algae index, the blue-green band ratio and the surface reflectance data into the U-Net convolutional neural network for segmentation processing to obtain the algal bloom coverage area.

[0015] The calculation module is used to perform integral processing on the particle size volume concentration distribution data to obtain the biological volume concentration at each depth layer, perform interpolation processing on the biological volume concentration at each depth layer to obtain the vertical profile distribution curve, and calculate the surface enrichment and floating trend index of the vertical profile distribution curve.

[0016] The registration module is used to extract the normalized algal index and the biological volume concentration of each depth layer at the corresponding location of the algal bloom coverage area to form a registration sample pair. The registration sample pair is trained by random forest regression, and the correlation between the normalized algal index and the biological volume concentration of each depth layer is updated based on the training results.

[0017] The determination module is used to input the normalized algal index, the surface enrichment degree, and the rising trend index into a CNN-LSTM model to obtain the probability of algal bloom, and to determine the warning level based on the algal bloom probability and the rising trend index.

[0018] Thirdly, an intelligent monitoring device for cyanobacterial blooms is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the intelligent monitoring device for cyanobacterial blooms to execute the aforementioned intelligent monitoring method for cyanobacterial blooms.

[0019] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored in the computer and, when executed on a computer, cause the computer to perform the aforementioned intelligent monitoring method for cyanobacterial blooms.

[0020] The technical solution provided in this application obtains surface reflectance data by collecting surface spectral reflectance data and underwater particle size distribution data and performing atmospheric correction processing. This overcomes the shortcomings of traditional monitoring methods that rely on only a single data source, resulting in incomplete information, and achieves simultaneous acquisition of surface and underwater monitoring information. Based on the surface reflectance data, the normalized algal index and blue-green band ratio are calculated. The spectral contrast between cyanobacteria and water is enhanced by utilizing the spectral response characteristics of cyanobacteria in different bands. These spectral indices and surface reflectance data are input into a U-Net convolutional neural network for segmentation. Through an encoder-decoder architecture and skip connection mechanism, multi-scale features are fused to accurately identify the spatial distribution of cyanobacteria on the water surface and obtain the bloom coverage area. Compared with traditional threshold segmentation methods, this method has higher recognition accuracy and robustness. The particle size distribution data is integrated to obtain the biomass concentration at each depth layer, and interpolation is used to obtain the vertical profile distribution curve. Then, the surface enrichment and buoyancy trend index are calculated. These characteristic parameters quantitatively characterize the vertical distribution and migration dynamics of cyanobacteria, providing key indicators for identifying early signs of blooms. By extracting the normalized algal index corresponding to the algal bloom coverage area and constructing registration sample pairs with the biomass concentration at each depth layer, and training with random forest regression, a quantitative correlation between surface spectral features and underwater biomass distribution was established. The random forest algorithm, by constructing multiple decision trees and calculating feature importance ranking, verified the predictive ability of the surface spectral index for underwater biomass, solving the problem of the inability to mutually verify and enhance each other due to the separation of surface and underwater data. The normalized algal index, surface enrichment, and buoyancy trend index were input into a CNN-LSTM model as time-series features. The CNN convolutional layer extracted local temporal pattern features to capture short-term change patterns, while the LSTM layer learned long-term temporal dependencies through memory units and gating mechanisms. This enabled the identification of the dynamic evolution process of cyanobacteria migrating and accumulating from underwater to the surface. After obtaining the algal bloom probability, the warning level was determined based on the probability threshold and the buoyancy trend index, realizing intelligent early warning decision-making based on multi-dimensional feature fusion. Compared with early warning methods based on a single data source, this method has higher accuracy and a longer lead time, providing a valuable time window for emergency response.

[0021] This application utilizes the U-Net convolutional neural network algorithm in the field of cyanobacterial bloom monitoring, fully leveraging its advantages in image semantic segmentation. By extracting multi-scale features through the encoder and restoring spatial resolution through upsampling by the decoder, pixel-level accurate identification of cyanobacterial distribution on the water surface is achieved. Compared with traditional spectral thresholding methods, it can better handle interference from complex aquatic environments and changes in illumination. A surface-underwater correlation model is established using the random forest regression algorithm, utilizing its strong fitting ability for nonlinear relationships and interpretability of feature importance. An ensemble learning mechanism reduces the overfitting risk of a single model, enabling stable inversion of underwater biomass distribution from surface spectral features, providing an effective technical means for the intelligent fusion of surface-underwater information. A CNN-LSTM fusion model was applied to predict algal blooms, combining the advantages of convolutional neural networks in spatial feature extraction with the advantages of long short-term memory networks in temporal modeling. The CNN convolutional layers capture local change patterns between adjacent time steps through one-dimensional convolution operations, while the LSTM layers selectively retain and forget information through gating mechanisms. This model can simultaneously model short-term fluctuations and long-term trends, effectively capturing the temporal evolution of cyanobacteria concentration and dynamic features such as surface enrichment and upward migration. Compared with traditional statistical prediction methods, it has stronger nonlinear fitting capabilities and higher prediction accuracy, enabling early prediction of the timing and intensity of algal blooms. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of one embodiment of the intelligent monitoring method for cyanobacterial blooms in this application.

[0024] Figure 2 This is a schematic diagram illustrating the correlation between the normalized algal index and the surface biomass concentration in the embodiments of this application;

[0025] Figure 3 This is a schematic diagram of one embodiment of the intelligent monitoring system for cyanobacterial blooms in this application.

[0026] Figure 4 This is a schematic block diagram of the intelligent monitoring device for cyanobacterial blooms in an embodiment of the present invention. Detailed Implementation

[0027] This application provides an intelligent monitoring method and system for cyanobacterial blooms. The terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0028] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent monitoring method for cyanobacterial blooms in this application includes:

[0029] Step S1: Collect water surface spectral reflectance data and underwater particle size volume concentration distribution data, and perform atmospheric correction processing on the water surface spectral reflectance data to obtain surface reflectance data;

[0030] Step S2: Calculate the normalized algal index and blue-green band ratio based on surface reflectance data. Input the normalized algal index, blue-green band ratio and surface reflectance data into the U-Net convolutional neural network for segmentation processing to obtain the algal bloom coverage area.

[0031] Step S3: Integrate the particle size volume concentration distribution data to obtain the biological volume concentration at each depth layer, interpolate the biological volume concentration at each depth layer to obtain the vertical profile distribution curve, and calculate the surface enrichment and floating trend index of the vertical profile distribution curve.

[0032] Step S4: Extract the normalized algal index corresponding to the algal bloom coverage area and the biological volume concentration of each depth layer to form a registration sample pair. Train the registration sample pair with random forest regression to obtain the surface-underwater correlation model.

[0033] Step S5: Input the normalized algal index, surface enrichment and rising trend index into the CNN-LSTM model to obtain the probability of algal bloom, and determine the warning level based on the algal bloom probability and rising trend index.

[0034] It is understood that the executing entity of this application can be an intelligent monitoring system for cyanobacterial blooms, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiments use a server as an example for illustration.

[0035] Specifically, after acquiring raw digital grayscale values ​​in the blue, green, red, and near-infrared bands using a multispectral camera, the grayscale values ​​were converted into radiance values ​​through radiometric calibration. The radiance values ​​were corrected and converted based on the known reflectivity of a standard reflector, and the water surface spectral reflectivity data was calculated using the Earth-Sun distance correction coefficient and the solar zenith angle. The surface reflectivity data was obtained by performing atmospheric scattering elimination processing on the water surface spectral reflectivity data using the dark pixel method. The dark pixel method assumes that the reflectivity of clean water in the near-infrared band is close to zero, and the reflectivity observed in this band is considered as a contribution from atmospheric scattering and subtracted from each band. Scattering intensity data from 32 particle size channels were acquired at four depths of 0.5 m, 1.0 m, 2.0 m, and 3.0 m using a LISST laser scattering sensor. The scattering intensity data were processed using the three-sigma criterion to remove outliers exceeding the mean ± 3 standard deviations. For example, if the average scattering intensity data at a water depth of 0.5 meters is 5.2 mg / L and the standard deviation is 0.8 mg / L, then the upper limit is 7.6 mg / L. Sampling points with values ​​of 9.1 mg / L exceed the upper limit and are discarded. The valid data after the rejection constitute the underwater particle size volume concentration distribution data.

[0036] Near-infrared and red-band reflectance are extracted from surface reflectance data. The normalized algae index is calculated by subtracting the red-band reflectance from the near-infrared reflectance and then dividing by the sum of the two. Blue and green-band reflectance are extracted, and the blue-green band ratio is obtained by dividing the blue-band reflectance by the green-band reflectance. The normalized algae index, the blue-green band ratio, and the four bands of surface reflectance data are stacked to form a 7-channel input tensor. This 7-channel input tensor is fed into the encoder of a U-Net convolutional neural network for feature extraction. The encoder outputs feature maps at different scales through convolutional and pooling layers. The multi-scale feature maps are upsampled and fused using skip connections by the decoder. The decoder outputs a feature map of the same size as the input image, which is then converted into a probability distribution for each pixel belonging to one of four categories: high-concentration algal bloom, medium-concentration algal bloom, low-concentration algal bloom, or clean water body, using a softmax activation function. After counting the number of pixels with high-concentration and medium-concentration algal blooms, multiply the number by the actual area corresponding to each pixel to obtain the algal bloom coverage area. For example, if the image size is 2000×2000 pixels, the spatial resolution is 0.5 meters, and the area of ​​each pixel is 0.25 square meters, then the count shows 150,000 pixels with high-concentration algal blooms and 80,000 pixels with medium-concentration algal blooms, resulting in an algal bloom coverage area of ​​230,000 pixels × 0.25 square meters = 57,500 square meters.

[0037] The total biomass concentration at each depth layer is obtained by multiplying and summing the volume concentration of each particle size channel in the underwater particle size distribution data with the corresponding channel width. The proportion of volume concentration in the 10-80 micrometer range is determined; when this proportion exceeds 70%, the depth layer is identified as the dominant cyanobacteria layer. The total biomass concentration at each depth layer is calculated by converting the total biomass concentration based on the average single-cell volume and the population correction factor. Cubic spline interpolation is performed on the biomass concentration at each depth layer to obtain continuous vertical profile distribution curves. The surface biomass concentration is calculated as the ratio of the surface biomass concentration to the average biomass concentration across multiple depth layers to obtain the surface enrichment. The buoyancy trend index is obtained by dividing the difference between the current surface biomass concentration and the concentration from the previous hour by the concentration from the previous hour. For example, the total biomass concentration at a depth of 0.5 meters, after multiplying and summing the 32 particle size channels, is 65.4 mg / L. The sum of the biomass concentrations within the 10-80 micrometer particle size range is 48.2 mg / L, accounting for 74%, exceeding the 70% threshold, thus classifying it as a cyanobacteria-dominant water layer. With an average single-cell volume of 20 cubic micrometers and a population correction factor of 0.8, the biomass concentration at a depth of 0.5 meters is 65.4 × 0.8 = 52.3 mg / L.

[0038] The geographic coordinates of the LISST laser scattering sensor were obtained. Using these coordinates as the center, all pixels within a 50-meter radius buffer zone were extracted from the distribution map corresponding to the algal bloom coverage area. The mean normalized algal index (NAL) of all pixels within the buffer zone was calculated, along with the mean blue-green band ratio of all pixels. The mean NAL and mean blue-green band ratio of the regions were then matched with the corresponding biomass concentrations at each depth level using timestamp matching to obtain registered sample pairs. These registered sample pairs were then used to train a random forest regression algorithm for decision tree construction and feature importance ranking. The random forest algorithm constructed multiple regression decision trees. Each tree recursively partitioned the feature space, and each non-leaf node selected a feature and its threshold for binary splitting to minimize the variance of the target variable in the child nodes. Feature importance ranking was obtained by summing the variance reduction brought by each feature in all trees, thus verifying the effectiveness of the association between the NAL and biomass concentrations at each depth level. For example, a sensor with coordinates of 102.5 degrees east longitude and 25.3 degrees north latitude has an extraction radius of 50 meters containing 320 pixels. The average normalized algae index of the region is 0.58, and the average blue-green band ratio of the region is 0.51. This, along with the biological volume concentration of the four depth layers, forms a registration sample pair.

[0039] A sliding time window was used to extract the normalized algal bloom index, surface enrichment, and rising trend index for the past six consecutive hours. Sampling was performed hourly to obtain feature data for six time steps. These feature data were then arranged chronologically to form a time-series feature. The time-series features were input into the convolutional layers of a CNN-LSTM model for one-dimensional convolution processing to extract local temporal pattern features. The convolutional layers used multiple convolutional kernels to slide across the time series to extract features. The feature tensors output from the convolutional layers were input into an LSTM layer to learn temporal dependencies, resulting in a high-dimensional feature vector. The LSTM layer includes memory units and gating mechanisms to capture long-term temporal dependencies. The high-dimensional feature vector was then mapped using a fully connected layer and activated with sigmoid to obtain the algal bloom probability, with a probability value between 0 and 1. The system determined whether the algal bloom probability exceeded a first threshold and whether the rising trend index exceeded a second threshold, and based on the results, a warning level was determined. For example, a blue alert is triggered when the probability of an algal bloom is ≥0.6 and the upward trend index is ≥0.15; a yellow alert is triggered when the probability of an algal bloom is ≥0.75 and the upward trend index is ≥0.25; and a red alert is triggered when the probability of an algal bloom is ≥0.85 and the upward trend index continues to rise. In a certain monitoring session, the normalized algal indexes extracted over the past 6 hours were 0.52, 0.54, 0.56, 0.58, 0.61, and 0.63; the surface enrichment was 1.25, 1.28, 1.32, 1.35, 1.38, and 1.42; and the upward trend indices were 0.08, 0.10, 0.12, 0.15, 0.18, and 0.21. Inputting these values ​​into a CNN-LSTM model yielded an algal bloom probability of 0.82. Combined with the upward trend index of 0.21 exceeding the 0.25 threshold, a yellow alert was triggered.

[0040] In one specific embodiment, step S1 includes:

[0041] The raw digital grayscale values ​​of blue, green, red and near-infrared bands are acquired by a multispectral camera, and the radiance values ​​are obtained by radiometric calibration of the raw digital grayscale values.

[0042] The radiance value is calibrated and converted based on the known reflectivity of the standard reflector, and the spectral reflectivity data of the water surface is calculated according to the sun-earth distance correction coefficient and the solar zenith angle.

[0043] The dark pixel method was used to perform atmospheric scattering elimination processing on the water surface spectral reflectance data to obtain the surface reflectance data;

[0044] The scattering intensity data of multiple particle size channels were collected at multiple depth layers using the LISST laser scattering sensor. The scattering intensity data were processed using the three sigma criterion to remove outliers that exceeded the mean standard deviation range, thus obtaining underwater particle size volume concentration distribution data.

[0045] Specifically, after acquiring raw digital grayscale values ​​in the blue, green, red, and near-infrared bands using a multispectral camera, radiometric calibration is performed to convert these grayscale values ​​into radiance values. Radiometric calibration is the process of converting the digitally quantized values ​​detected by the sensor into physical radiation quantities, achieved through a linear transformation using pre-calibrated gain coefficients and offsets. The radiance values ​​are then radiometrically calibrated based on the known reflectivity of a standard reflector, which is a reference target with uniform surface reflectivity and accurately measured reflectivity; typically, a gray board with 50% reflectivity or a white board with 99% reflectivity is used as the calibration standard. The water surface spectral reflectivity data is calculated based on the Earth-Sun distance correction factor and the solar zenith angle. The Earth-Sun distance correction factor is used to correct for seasonal variations in irradiance caused by the Earth's revolution around the Sun, and the solar zenith angle is the angle between sunlight and the ground normal, affecting the path length of solar radiation transmission in the atmosphere. Surface reflectance data were obtained by performing atmospheric scattering elimination processing on water surface spectral reflectance data using the dark pixel method. The dark pixel method assumes that the reflectance of clean water in the near-infrared band is close to zero, and the reflectance observed in this band is considered as a contribution from atmospheric scattering, with the corresponding atmospheric scattering component subtracted from the reflectance of each band. Scattering intensity data for 32 particle size channels were collected at four depths (0.5 m, 1.0 m, 2.0 m, and 3.0 m) using a LISST laser scattering sensor. Each particle size channel corresponds to a specific particle size range, and the scattering intensity reflects the volume concentration of particles within that range. The scattering intensity data were processed using the three-sigma criterion to remove outliers exceeding the mean ± 3 standard deviations. The three-sigma criterion is based on the normal distribution theory, and the standard deviation is obtained by averaging the sum of the squares of the differences between each data point and the mean, and then taking the square root. For example, if the mean scattering intensity data at a water depth of 0.5 meters is 5.2 mg / L and the standard deviation is 0.8 mg / L, then the upper limit is 7.6 mg / L and the lower limit is 2.8 mg / L. Sampling points with a value of 9.1 mg / L exceed the upper limit and are discarded. The valid data after the rejection constitute the underwater particle size volume concentration distribution data.

[0046] In one specific embodiment, step S2 includes:

[0047] The near-infrared and red light reflectance were extracted from the surface reflectance data, and the normalized algal index was calculated based on the ratio of the difference between the near-infrared and red light reflectance.

[0048] Extract the blue and green light reflectance from the surface reflectance data, and calculate the blue-green band ratio based on the ratio of blue light reflectance to green light reflectance.

[0049] The normalized algae index, the blue-green band ratio, and the four bands of surface reflectance data are combined to form a seven-channel input tensor. The seven-channel input tensor is then input into the encoder of the U-Net convolutional neural network for feature extraction to obtain a multi-scale feature map.

[0050] The multi-scale feature map is upsampled by the decoder and fused by skip connections to output the classification probability of cyanobacteria concentration for each pixel. The coverage area of ​​the cyanobacteria bloom is obtained by counting the number of pixels with high concentration and medium concentration.

[0051] Specifically, after extracting the near-infrared and red light reflectance from the surface reflectance data, the normalized algal index is calculated by subtracting the red light reflectance from the near-infrared reflectance, and then dividing the difference by the sum of the near-infrared and red light reflectance. The normalized algal index utilizes the spectral characteristics of cyanobacteria—high reflectance in the near-infrared band and low reflectance in the red light band—to eliminate the influence of light intensity variations through normalization, thus enhancing the spectral contrast between cyanobacteria and the water body. After extracting the blue and green light reflectance from the surface reflectance data, the blue-green ratio is directly obtained by dividing the blue light reflectance by the green light reflectance. The blue-green ratio reflects the relative content of phycocyanin in the water body. Phycocyanin is a unique accessory pigment of cyanobacteria that has strong absorption in the blue light band, leading to a decrease in blue light reflectance. The normalized algae index, blue-green band ratio, and surface reflectance data (four bands: blue, green, red, and near-infrared) are stacked according to channel dimension to form a 7-channel input tensor. This 7-channel input tensor is then fed into the encoder of a U-Net convolutional neural network for feature extraction. The encoder contains multiple convolutional and pooling layers. The convolutional layers extract features from the local receptive field of the input tensor using sliding convolution kernels, while the pooling layers downsample the output of the convolutional layers to reduce spatial resolution. The encoder outputs feature maps at different scales; shallow feature maps retain spatial details, while deeper feature maps contain semantic abstraction. The multi-scale feature maps are then upsampled and fused using skip connections. The decoder upsamples the feature maps using transposed convolutions to restore spatial resolution, and skip connections concatenate the feature maps of corresponding layers in the encoder with those of the current layer in the decoder, fusing feature information from different levels. The decoder finally outputs a feature map of the same size as the input image, which is then converted using a softmax activation function to a probability distribution for each pixel belonging to one of four categories: high-concentration algal bloom, medium-concentration algal bloom, low-concentration algal bloom, or clean water. After counting the number of pixels with high-concentration and medium-concentration algal blooms, the algal bloom coverage area is obtained by multiplying the number of pixels by the actual area corresponding to a single pixel. For example, if the image size is 2000×2000 pixels, the spatial resolution is 0.5 meters, the area of ​​a single pixel is 0.25 square meters, and the near-infrared reflectance of a pixel is 0.45 and the red light reflectance is 0.12, the normalized algal index is calculated as (0.45-0.12) / (0.45+0.12)=0.58, the blue light reflectance is 0.08 and the green light reflectance is 0.15, and the blue-green ratio is calculated as 0.08 / 0.15=0.53. The statistics show 150,000 pixels with high-concentration algal blooms and 80,000 pixels with medium-concentration algal blooms, with an algal bloom coverage area of ​​230,000 × 0.25 = 57,500 square meters.

[0052] In one specific embodiment, step S3 includes:

[0053] The volume concentration of each particle size channel in the underwater particle size volume concentration distribution data is multiplied and accumulated with the corresponding particle size channel width to obtain the total biological volume concentration of each depth layer.

[0054] The proportion of the volume concentration of particles within the set range to the total biological volume concentration is determined. When the proportion exceeds the preset threshold, the depth layer is determined to be the dominant cyanobacteria water layer. The biological volume concentration of each depth layer is calculated by converting the total biological volume concentration based on the average single cell volume and the population correction factor.

[0055] Cubic spline interpolation was performed on the biological volume concentration at each depth layer to obtain continuous vertical profile distribution curves.

[0056] The surface enrichment is obtained by calculating the ratio of the surface biological volume concentration to the average biological volume concentration of multiple depth layers. The rising trend index is obtained by calculating the ratio of the difference between the current surface biological volume concentration and the surface biological volume concentration of the previous hour to the surface biological volume concentration of the previous hour.

[0057] Specifically, the total biomass concentration at each depth is obtained by multiplying and summing the volume concentration of each particle size channel in the underwater particle size distribution data with the corresponding channel width. The channel width is the span of the particle size range covered by that channel. The proportion of the volume concentration of particles within a set range (typically 10 to 80 micrometers) to the total biomass concentration is determined. This range represents the typical particle size range for single-celled cyanobacteria and small colonies. When this proportion exceeds a preset threshold, such as 70%, the depth is identified as the dominant cyanobacteria layer. The total biomass concentration at each depth is calculated by converting the total biomass concentration based on the average single-cell volume and a colony correction factor. The average single-cell volume is obtained by averaging the volumes of a large number of cyanobacteria cells observed under a microscope. The colony correction factor is used to correct the deviation of laser scattering measurements when cyanobacteria exist in colony form, as the intercellular spaces within the colony affect the scattering characteristics. Cubic spline interpolation was performed on the biomass concentration at various depths to obtain continuous vertical profile distribution curves. Cubic spline interpolation is a smooth interpolation method that constructs a cubic polynomial function between known discrete data points, ensuring the interpolation curve is continuous at the data points, and that the first and second derivatives are also continuous, avoiding the broken line effect caused by piecewise linear interpolation. The ratio of the surface biomass concentration to the average biomass concentration across multiple depths was calculated to obtain the surface enrichment degree. A surface enrichment degree greater than 1 indicates that cyanobacteria are accumulating at the surface, and the higher the value, the higher the enrichment degree. The difference between the current surface biomass concentration and the concentration one hour prior was divided by the concentration one hour prior to obtain the upward trend index. A positive upward trend index indicates an increase in surface biomass, and the higher the value, the faster the growth rate, indicating that cyanobacteria are migrating from deeper layers to the surface. For example, after multiplying and summing the 32 particle size channels at a depth of 0.5 meters, the total biomass concentration is calculated to be 65.4 mg / L. The sum of the volume concentrations of all channels within the particle size range of 10-80 micrometers is 48.2 mg / L, accounting for 48.2 / 65.4 = 74% of the total biomass concentration. This exceeds the 70% threshold and is considered a cyanobacteria-dominant water layer. With an average single-cell volume of 20 cubic micrometers and a population correction factor of 0.8, the biomass concentration at a depth of 0.5 meters is 65.4 × 0.8 = 52.3 mg / L. Similarly, the concentrations at depths of 1.0 meters are 45.6 mg / L, 2.0 meters are 32.1 mg / L, and 3.0 meters are 21.8 mg / L. Vertical profile distribution curves were constructed by cubic spline interpolation of data from four depth layers. The average biological volume concentration was calculated to be (52.3+45.6+32.1+21.8) / 4=37.95 mg / L, the surface enrichment was 52.3 / 37.95=1.38, the surface biological volume concentration in the first hour was 46.7 mg / L, and the rising trend index was (52.3-46.7) / 46.7=0.12.

[0058] In one specific embodiment, step S4 includes:

[0059] Obtain the geographic coordinates of the LISST laser scattering sensor, and extract all pixels within a buffer zone of a set radius from the distribution map corresponding to the algal bloom coverage area, centered on the geographic coordinates.

[0060] The mean value of the normalized algae index of the region is obtained by averaging the normalized algae index of all pixels within the buffer range, and the mean value of the blue-green band ratio of the region is obtained by averaging the blue-green band ratio of all pixels within the buffer range.

[0061] The normalized algal index mean and the blue-green band ratio mean of the region are matched with the biological volume concentration of each depth layer at the corresponding time to obtain the registered sample pairs.

[0062] The registered samples were used to train the random forest regression algorithm to construct decision trees and rank features by importance, and the effectiveness of the association between the normalized algal index and the biological volume concentration at each depth layer was verified.

[0063] Specifically, after obtaining the geographic coordinates of the LISST laser scattering sensor (including longitude and latitude), all pixels within a buffer zone with a set radius (e.g., 50 meters) are extracted from the distribution map corresponding to the algal bloom coverage area, centered on these coordinates. The buffer zone is a circular area with a set radius centered at a point, and pixels within this zone are considered to be spatially correlated with the sensor's location. The mean normalized algal bloom index (NUP) of all pixels within the buffer zone is calculated by averaging the NUP values ​​of all pixels within the buffer zone and dividing by the total number of pixels. The mean blue-green band ratio of all pixels within the buffer zone is also calculated by averaging the NUP values ​​of all pixels within the buffer zone, using the same method as for the mean NUP. Registered sample pairs are obtained by timestamping the mean normalized algal index and the mean blue-green band ratio of the region with the corresponding biomass concentrations at each depth. Timestamping involves pairing surface data with underwater data collected at the closest times, typically requiring a time difference of less than a certain threshold, such as 8 minutes, because cyanobacteria distribution is relatively stable over short periods. The registered sample pairs are then input into a random forest regression algorithm for decision tree construction and feature importance ranking training. The random forest regression algorithm, based on ensemble learning, constructs multiple regression decision trees, each randomly sampling a subset of samples and features during training. The decision trees construct their structure by recursively partitioning the feature space. Each non-leaf node selects a feature and its threshold for binary splitting, minimizing the variance of the target variable in the child nodes. Leaf nodes store the average target variable of the samples in that region as the predicted value. After training, for newly input surface spectral features, each tree in the random forest independently makes a prediction, and the final prediction is the average of all tree predictions. Feature importance ranking is obtained by summing the variance reduction brought by each feature when used for partitioning across all trees. The higher the importance value, the greater the contribution of the feature to the prediction target, thus verifying the effectiveness of the correlation between the normalized algal index and the biomass concentration at each depth layer. For example, a LISST sensor with geographical coordinates of 102.5 degrees east longitude and 25.3 degrees north latitude is used to extract a buffer zone with a radius of 50 meters centered on this coordinate in the algal bloom coverage distribution map. This buffer zone contains 320 pixels. The normalized algal index is extracted by traversing these 320 pixels. The sum of the 320 values ​​is 185.6. Dividing by 320, the mean normalized algal index of the region is 0.58. Similarly, the mean blue-green band ratio of the region is calculated to be 0.51. The sensor collected data at 10:30 AM, and the corresponding biological volume concentrations at each depth were 52.3 mg / L at 0.5 m, 45.6 mg / L at 1.0 m, 32.1 mg / L at 2.0 m, and 21.8 mg / L at 3.0 m. This set of data constitutes one registration sample pair.

[0064] Figure 2This is a schematic diagram illustrating the correlation between the normalized algal index and the surface biomass concentration in an embodiment of this application. Figure 2 As shown, there is a significant positive correlation between the normalized algal index and the surface biomass concentration. The horizontal axis represents the normalized algal index, ranging from 0.45 to 0.63, and the vertical axis represents the surface biomass concentration in milligrams per liter, ranging from 35.2 to 60.2. The data points on the curve correspond to the registered sample pairs collected in actual monitoring. The continuity of the curve reflects the fitting relationship obtained through random forest regression training. As the normalized algal index increases, the surface biomass concentration shows a stable upward trend, verifying the predictive ability of water surface spectral characteristics for underwater biomass and providing a quantitative basis for achieving water surface-underwater data fusion. The example data points shown in the tooltip (normalized algal index 0.45 corresponds to a biomass concentration of 35.2 milligrams per liter) reflect the model's correlation accuracy in specific values. The establishment of this correlation solves the technical problem of the lack of effective fusion between water surface spectral data and underwater vertical profile data in existing technologies.

[0065] In one specific embodiment, step S5 includes:

[0066] A sliding time window was used to extract the normalized algal index, surface enrichment, and rising trend index from multiple consecutive hours in the past, and the extracted features from multiple time steps were combined to form a time series feature.

[0067] The time series features are input into the convolutional layer of the CNN-LSTM model for one-dimensional convolution processing to extract local time pattern features. The local time pattern features are then activated and discarded to obtain the convolutional feature tensor.

[0068] The convolutional feature tensor is input into the LSTM layer to learn the temporal dependency relationship and obtain a high-dimensional feature vector. The high-dimensional feature vector is then mapped by a fully connected layer and activated by sigmoid to obtain the probability of algal bloom.

[0069] Determine whether the probability of an algal bloom exceeds the first threshold and whether the upward trend index exceeds the second threshold. Based on the determination results, the warning level is set as blue, yellow, or red.

[0070] Specifically, a sliding time window is used to extract the normalized algae index, surface enrichment, and rising trend index from the past six consecutive hours. The sliding time window is a method of extracting time-series segments from historical data with a fixed time span, sampling once per hour to obtain feature data for six time steps. The extracted features from multiple time steps are arranged chronologically to form a time-series feature matrix, with each time step containing three feature variables, constituting a 6×3 dimensional feature matrix. The time-series features are then input into the convolutional layers of a CNN-LSTM model for one-dimensional convolution processing to extract local temporal pattern features. The convolutional layers use multiple convolutional kernels that slide across the time series; the kernel size is typically 3, meaning that data from three consecutive time steps are processed each time, capturing local change patterns between adjacent time steps through convolution operations. The local temporal pattern features are then activated and discarded to obtain a convolutional feature tensor. Activation uses the ReLU activation function to set negative values ​​to zero and retain positive values. Discarding uses a dropout mechanism to randomly set the outputs of some neurons to zero; the dropout rate is typically set to 0.3 to prevent model overfitting. The convolutional feature tensor is input into an LSTM layer to learn temporal dependencies, resulting in a high-dimensional feature vector. The LSTM layer contains memory units, input gates, forget gates, and output gates. Memory units store long-term information, input gates control the writing of new information, forget gates control the discarding of old information, and output gates control the reading of information. These gating mechanisms capture long-term temporal dependencies. The high-dimensional feature vector is then mapped using a fully connected layer and activated by a sigmoid function to obtain the algal bloom probability. The fully connected layer maps the high-dimensional feature vector to a single numerical value, and the sigmoid activation function compresses this value to between 0 and 1 to represent the probability. The system determines whether the algal bloom probability exceeds a first threshold and whether the upward trend index exceeds a second threshold. Based on the determination, a warning level is established: a blue warning is triggered when the algal bloom probability is ≥0.6 and the upward trend index is ≥0.15; a yellow warning is triggered when the algal bloom probability is ≥0.75 and the upward trend index is ≥0.25; and a red warning is triggered when the algal bloom probability is ≥0.85 and the upward trend index continues to rise. For example, in a certain monitoring, the normalized algal index extracted over the past 6 hours was 0.52, 0.54, 0.56, 0.58, 0.61, and 0.63, respectively; the surface enrichment was 1.25, 1.28, 1.32, 1.35, 1.38, and 1.42, respectively; and the rising trend index was 0.08, 0.10, 0.12, 0.15, 0.18, and 0.21, respectively. These data constituted a 6×3 time series feature input to a CNN-LSTM model. The model output an algal bloom probability of 0.82. Combined with the rising trend index of 0.21 exceeding the threshold of 0.25, a yellow warning was triggered.

[0071] In one specific embodiment, a sliding time window is used to extract the normalized algal index, surface enrichment, and rising trend index from multiple consecutive hours over the past. The extracted features from multiple time steps are combined to form a time series feature, including:

[0072] The sliding time window length was set to a continuous 6 hours. Normalized algae index, surface enrichment and rising trend index were extracted from historical monitoring data according to the hourly sampling frequency to obtain feature data for 6 time steps.

[0073] The normalized algae index, surface enrichment and rising trend index of each time step in the feature data of the 6 time steps were numerically normalized to obtain the normalized time step feature values.

[0074] The normalized time step feature values ​​are arranged in chronological order to form a feature matrix. The row dimension of the feature matrix corresponds to the number of time steps, and the column dimension corresponds to the number of feature variables.

[0075] Tensor transformation is performed on the feature matrix to obtain time series features with a shape equal to the time step length multiplied by the feature dimension.

[0076] Specifically, a sliding time window of 6 consecutive hours was set. Normalized algae index, surface enrichment, and rising trend index were extracted from historical monitoring data at an hourly sampling frequency, resulting in feature data for 6 time steps. The sliding time window is a time series data processing method that extracts data segments from a continuous monitoring data stream over a fixed time span. An hourly sampling frequency means data is recorded once every hour. The normalized algae index, surface enrichment, and rising trend index for each of the 6 time steps were numerically normalized to obtain normalized feature values ​​for that time step. Numerical normalization scales features with different dimensions and numerical ranges to a uniform numerical interval. A minimum-maximum normalization method is typically used, calculated as the normalized value equal to the original value minus the minimum value, divided by the maximum value minus the minimum value. This maps the value to between 0 and 1, eliminating the impact of dimensional differences between features on model training. The normalized time-step feature values ​​are arranged in chronological order to form a feature matrix. The row dimension of the feature matrix corresponds to the number of time steps, and the column dimension corresponds to the number of feature variables. The feature matrix is ​​a two-dimensional array; rows represent the time dimension, and columns represent the feature dimension. Each element in the matrix represents a specific feature value at a specific time step, and the chronological order ensures the sequential relationship of the time series. Tensor transformation is then performed on the feature matrix to obtain a time series feature matrix with the shape of the time step length multiplied by the feature dimension. A tensor is a mathematical expression of a multi-dimensional array; tensor transformation encapsulates the feature matrix into a tensor format that can be processed by deep learning frameworks, with a 6×3 shape representing 6 time steps and 3 feature variables. For example, in a certain monitoring, the normalized algal index extracted over the past 6 hours was 0.52, 0.54, 0.56, 0.58, 0.61, and 0.63, respectively; the surface enrichment was 1.25, 1.28, 1.32, 1.35, 1.38, and 1.42, respectively; and the rising trend index was 0.08, 0.10, 0.12, 0.15, 0.18, and 0.21, respectively. The normalized algal index was normalized, with a minimum value of 0.52 and a maximum value of 0.63. The normalized value for the first time step was (0.52-0.52) / (0.63-0.52)=0, and the normalized value for the sixth time step was (0.63-0.52) / (0.63-0.52)=1. The values ​​for other time steps were calculated sequentially. After normalizing the surface enrichment and upward trend index, the normalized data are arranged in chronological order. The first row contains the three feature values ​​of the first time step, and the sixth row contains the three feature values ​​of the sixth time step, forming a 6×3 feature matrix. This matrix is ​​then converted into a 6×3 tensor as the time series feature.

[0077] The above describes the intelligent monitoring method for cyanobacterial blooms in the embodiments of this application. The following describes the intelligent monitoring system for cyanobacterial blooms in the embodiments of this application. Please refer to [link / reference]. Figure 3 One embodiment of the intelligent monitoring system for cyanobacterial blooms in this application includes:

[0078] The correction module is used to collect water surface spectral reflectance data and underwater particle size volume concentration distribution data, and to perform atmospheric correction processing on the water surface spectral reflectance data to obtain surface reflectance data.

[0079] The segmentation module is used to calculate the normalized algae index and the blue-green band ratio based on the surface reflectance data, and input the normalized algae index, the blue-green band ratio and the surface reflectance data into the U-Net convolutional neural network for segmentation processing to obtain the algal bloom coverage area.

[0080] The calculation module is used to perform integral processing on the particle size volume concentration distribution data to obtain the biological volume concentration at each depth layer, perform interpolation processing on the biological volume concentration at each depth layer to obtain the vertical profile distribution curve, and calculate the surface enrichment and floating trend index of the vertical profile distribution curve.

[0081] The registration module is used to extract the normalized algal index and the biological volume concentration of each depth layer at the corresponding location of the algal bloom coverage area to form a registration sample pair. The registration sample pair is trained by random forest regression, and the correlation between the normalized algal index and the biological volume concentration of each depth layer is updated based on the training results.

[0082] The determination module is used to input the normalized algal index, the surface enrichment degree, and the rising trend index into a CNN-LSTM model to obtain the probability of algal bloom, and to determine the warning level based on the algal bloom probability and the rising trend index.

[0083] above Figure 3 The intelligent monitoring system for cyanobacterial blooms in this embodiment of the invention is described in detail from the perspective of modular functional entities. The intelligent monitoring device for cyanobacterial blooms in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0084] Reference Figure 4 This invention also provides an intelligent monitoring device for cyanobacterial blooms. This intelligent monitoring device can be a server, and its internal structure can be as follows: Figure 4As shown, the intelligent monitoring device for cyanobacterial blooms includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor in this computer design provides computing and control capabilities. The memory of the intelligent monitoring device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the intelligent monitoring device for cyanobacterial blooms stores the data corresponding to this embodiment. The network interface of the intelligent monitoring device for cyanobacterial blooms is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0085] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the intelligent monitoring device for cyanobacterial blooms to which the present invention is applied.

[0086] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the intelligent monitoring method for cyanobacterial blooms.

[0087] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0088] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an intelligent monitoring device for cyanobacterial blooms (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0089] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart monitoring method for cyanobacterial blooms, characterized in that, The method includes: Step S1: Collect water surface spectral reflectance data and underwater particle size volume concentration distribution data, and perform atmospheric correction processing on the water surface spectral reflectance data to obtain surface reflectance data; Step S2: Calculate the normalized algal index and blue-green band ratio based on the surface reflectance data. Input the normalized algal index, the blue-green band ratio, and the surface reflectance data into a U-Net convolutional neural network for segmentation processing to obtain the algal bloom coverage area. Step S3: Integrate the particle size volume concentration distribution data to obtain the biological volume concentration at each depth layer, interpolate the biological volume concentration at each depth layer to obtain the vertical profile distribution curve, and calculate the surface enrichment and rising trend index of the vertical profile distribution curve. The surface enrichment is the ratio of the surface biological volume concentration to the average biological volume concentration at multiple depth layers, and the rising trend index is the ratio of the difference between the surface biological volume concentration at the current moment and the surface biological volume concentration at the previous moment to the surface biological volume concentration at the previous moment. Step S4: Extract the normalized algal index and the biological volume concentration of each depth layer corresponding to the algal bloom coverage area to form a registration sample pair. Train the registration sample pair using random forest regression and update the correlation between the normalized algal index and the biological volume concentration of each depth layer based on the training results. Step S5: Input the normalized algal index, the surface enrichment degree, and the rising trend index into the CNN-LSTM model to obtain the algal bloom probability, and determine the warning level based on the algal bloom probability and the rising trend index.

2. The intelligent monitoring method for cyanobacterial blooms according to claim 1, characterized in that, Step S1 includes: The original digital grayscale values ​​of blue light band, green light band, red light band and near-infrared band are acquired by a multispectral camera, and the original digital grayscale values ​​are radiometrically calibrated to obtain the radiance value. The radiance value is radiometrically calibrated based on the known reflectivity of a standard reflector, and the water surface spectral reflectivity data is calculated based on the Earth-Sun distance correction factor and the solar zenith angle. The surface reflectance data is obtained by performing atmospheric scattering elimination processing on the water surface spectral reflectance data using the dark pixel method. The underwater particle size volume concentration distribution data is obtained by collecting scattering intensity data of multiple particle size channels at multiple depth layers using the LISST laser scattering sensor and processing the scattering intensity data using the three sigma criterion to remove outliers that exceed the mean standard deviation range.

3. The intelligent monitoring method for cyanobacterial blooms according to claim 1, characterized in that, Step S2 includes: The near-infrared and red light band reflectances are extracted from the surface reflectance data, and the normalized algal index is calculated based on the ratio of the difference between the near-infrared and red light band reflectances to the sum of their values. Extract the blue light band reflectance and green light band reflectance from the surface reflectance data, and calculate the blue-green band ratio based on the ratio of the blue light band reflectance to the green light band reflectance. The normalized algae index, the blue-green band ratio, and the four bands of the surface reflectance data are combined to form a seven-channel input tensor. The seven-channel input tensor is then input into the encoder of the U-Net convolutional neural network to extract features and obtain a multi-scale feature map. The multi-scale feature map is subjected to decoder upsampling and skip connection fusion processing to output the classification probability of cyanobacteria concentration for each pixel. The coverage area of ​​the cyanobacteria bloom is obtained by counting the number of pixels with high concentration and medium concentration.

4. The intelligent monitoring method for cyanobacterial blooms according to claim 1, characterized in that, Step S3 includes: The volume concentration of each particle size channel in the underwater particle size volume concentration distribution data is multiplied and accumulated with the corresponding particle size channel width to obtain the total biological volume concentration of each depth layer. The proportion of the volume concentration of particles within a set range to the total biological volume concentration is determined. When the proportion exceeds a preset threshold, the depth layer is determined to be a cyanobacteria-dominant water layer. The biological volume concentration of each depth layer is obtained by converting the total biological volume concentration based on the average single cell volume and the population correction factor. The biological volume concentration at each depth layer of multiple depth layers is subjected to cubic spline interpolation to obtain a continuous vertical profile distribution curve. The surface enrichment is obtained by calculating the ratio of the surface biological volume concentration to the average biological volume concentration of multiple depth layers, and the rising trend index is obtained by calculating the ratio of the difference between the current surface biological volume concentration and the surface biological volume concentration of the previous hour to the surface biological volume concentration of the previous hour.

5. The intelligent monitoring method for cyanobacterial blooms according to claim 1, characterized in that, Step S4 includes: Obtain the geographic coordinates of the LISST laser scattering sensor, and extract all pixels within a buffer zone of a set radius from the distribution map corresponding to the algal bloom coverage area, using the geographic coordinates as the center. The mean value of the normalized algae index of the region is obtained by averaging the normalized algae index of all pixels within the buffer range, and the mean value of the blue-green band ratio of the region is obtained by averaging the blue-green band ratio of all pixels within the buffer range. The normalized algal index mean of the region, the blue-green band ratio mean of the region, and the biological volume concentration of each depth layer at the corresponding time are timestamped to obtain the registration sample pair. The registered samples were used to train a random forest regression algorithm to construct a decision tree and rank features by importance, thereby verifying the effectiveness of the correlation between the normalized algal index and the biological volume concentration at each depth layer.

6. The intelligent monitoring method for cyanobacterial blooms according to claim 1, characterized in that, Step S5 includes: The normalized algal index, surface enrichment, and rising trend index of the past several consecutive hours are extracted using a sliding time window, and the extracted features of multiple time steps are combined to form the time series features. The time series features are input into the convolutional layer of the CNN-LSTM model for one-dimensional convolution processing to extract local time pattern features. The local time pattern features are then activated and discarded to obtain the convolutional feature tensor. The convolutional feature tensor is input into an LSTM layer to learn temporal dependencies and obtain a high-dimensional feature vector. The high-dimensional feature vector is then mapped by a fully connected layer and activated by a sigmoid layer to obtain the algal bloom probability. Determine whether the probability of algal bloom exceeds a first threshold and whether the upward trend index exceeds a second threshold, and determine the warning level as blue, yellow or red based on the determination result.

7. The intelligent monitoring method for cyanobacterial blooms according to claim 6, characterized in that, The process involves extracting the normalized algal index, surface enrichment, and rising trend index from multiple consecutive hours using a sliding time window. The extracted features from multiple time steps are then combined to form the time series features, including: The sliding time window length is set to a continuous 6 hours. The normalized algae index, the surface enrichment and the rising trend index are extracted from historical monitoring data according to the hourly sampling frequency to obtain feature data for 6 time steps. The normalized algae index, surface enrichment and rising trend index of each time step in the feature data of the 6 time steps are numerically normalized to obtain the normalized time step feature values. The normalized time step feature values ​​are arranged in chronological order to form a feature matrix, where the row dimension of the feature matrix corresponds to the number of time steps and the column dimension corresponds to the number of feature variables. The feature matrix is ​​subjected to tensor transformation to obtain the time series feature with a shape that is the time step length multiplied by the feature dimension.

8. An intelligent monitoring system for cyanobacterial blooms, characterized in that, The intelligent monitoring system for cyanobacterial blooms, as described in any one of claims 1-7, comprises: The correction module is used to collect water surface spectral reflectance data and underwater particle size volume concentration distribution data, and to perform atmospheric correction processing on the water surface spectral reflectance data to obtain surface reflectance data. The segmentation module is used to calculate the normalized algae index and the blue-green band ratio based on the surface reflectance data, and input the normalized algae index, the blue-green band ratio and the surface reflectance data into the U-Net convolutional neural network for segmentation processing to obtain the algal bloom coverage area; The calculation module is used to perform integral processing on the particle size volume concentration distribution data to obtain the biological volume concentration at each depth layer, perform interpolation processing on the biological volume concentration at each depth layer to obtain the vertical profile distribution curve, and calculate the surface enrichment and floating trend index of the vertical profile distribution curve. The registration module is used to extract the normalized algal index and the biological volume concentration of each depth layer at the corresponding location of the algal bloom coverage area to form a registration sample pair. The registration sample pair is trained by random forest regression, and the correlation between the normalized algal index and the biological volume concentration of each depth layer is updated based on the training results. The determination module is used to input the normalized algal index, the surface enrichment degree, and the rising trend index into a CNN-LSTM model to obtain the probability of algal bloom, and to determine the warning level based on the algal bloom probability and the rising trend index.

9. An intelligent monitoring device for cyanobacterial blooms, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the intelligent monitoring method for cyanobacterial blooms as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it causes the processor to execute the intelligent monitoring method for cyanobacterial blooms as described in any one of claims 1 to 7.

Images