Automatic identification method and system of algal blooms based on multispectral information

Abstract

The present application relates to the technical field of automatic identification of algal blooms, and discloses a method and system for automatic identification of algal blooms based on multispectral information. The method first performs standardization processing and mosaic fusion on the original multispectral images to obtain high-quality reflectance data. Subsequently, the water body range is accurately defined by calculating the water-vegetation absorption index and performing intelligent extraction, eliminating non-water body interference. The potential algal bloom areas in the water body are adaptively screened using the band ratio index to narrow the classification range and improve efficiency. Finally, multidimensional feature vectors including spectral reflectance, band ratio index, yellowness index and reflection valley depth are extracted from these potential algal bloom areas, and are input into a pre-trained machine learning classifier model for deep learning and pattern recognition, thereby overcoming the limitations of traditional methods that rely only on a single or a few empirical indices, and being able to capture the more complex and subtle spectral characteristics of algal blooms, achieving high-precision classification and discrimination.

Description

Technical Field

[0001] The present application relates to the technical field of automatic identification of algal blooms, and more specifically, to a method and system for automatic identification of algal blooms based on multispectral information. Background Art

[0002] Algal blooms, also known as water blooms, are a typical manifestation of eutrophication. Their large-scale outbreaks not only consume dissolved oxygen in the water, leading to the death of a large number of aquatic organisms and disrupting the water ecological balance, but also pose a serious threat to fishery production, drinking water safety and even human health. Therefore, timely and accurate monitoring and early warning of algal blooms are of vital importance for water environment management and ecological protection. Traditional algal bloom monitoring methods mostly rely on on-site sampling and laboratory analysis. These methods are usually time-consuming and labor-intensive, making it difficult to achieve large-scale, high-frequency dynamic monitoring. In addition, there is a lag in data acquisition, which cannot meet the needs of rapid response and refined management.

[0003] With the development of remote sensing technology, particularly multispectral remote sensing monitoring, it has become an important tool for algal bloom monitoring due to its non-contact, large-scale, and high-efficiency characteristics. Currently, a variety of algal bloom detection methods based on spectral indices have been developed, such as the floating algae index (FAI), the band ratio index (IRB), and the normalized difference vegetation index (NDVI). These methods identify algal bloom areas by analyzing the differences in the reflectance characteristics of algal blooms across different spectral bands. However, these methods based on a single or a few spectral indices often rely on empirical thresholds or fixed discrimination rules, resulting in poor accuracy and robustness in complex water environments (such as turbid water, interference from aquatic vegetation, and large variations in algal bloom concentrations). For example, simple band ratios may not be able to effectively distinguish algal blooms from suspended sediment or other aquatic plants in the water, leading to false or missed detections. Furthermore, the spectral characteristics of algal blooms may vary across regions and seasons, making fixed thresholds difficult to adapt to changing environments and resulting in insufficient generalization of the algorithms. Therefore, in order to overcome the deficiencies of the prior art, the present application provides an automatic identification solution for algal blooms based on multispectral information. Summary of the Invention

[0004] In order to solve the above technical problems, the present application is proposed. The embodiments of the present application provide a method and system for automatic identification of algal blooms based on multispectral information, which can adaptively learn the complex nonlinear relationship between algal blooms and non-algal blooms, effectively overcoming the problem of poor adaptability of traditional fixed threshold methods in different regions and seasons, and significantly reducing the rates of false positives and missed positives.

[0005] According to one aspect of the present application, a method for automatic identification of algal blooms based on multispectral information is provided, comprising: performing standardized preprocessing and mosaic fusion on the acquired original multispectral image sequence to obtain a multi-band ground reflectance mosaic map; calculating the water-vegetation absorption index of each pixel position in the multi-band ground reflectance mosaic map, and performing intelligent water body extraction on the multi-band ground reflectance mosaic map based on the water-vegetation absorption index to obtain a binary water body range mask; extracting a multi-band ground reflectance water body ROI map from the multi-band ground reflectance mosaic map based on the binary water body range mask; calculating the multi-band ground reflectance water body ROI map. I, and based on the band ratio index of each pixel position in the multi-band ground reflectance water body ROI map, the multi-band ground reflectance ROI map is adaptively screened for potential algal bloom areas to obtain a rough mask of potential algal bloom areas; based on the rough mask of potential algal bloom areas, a multi-band ground reflectance potential algal bloom area ROI map is extracted from the multi-band ground reflectance mosaic map; based on the multi-dimensional feature vector of each pixel position in the multi-band ground reflectance potential algal bloom area ROI map and a pre-trained machine learning classifier model, a binary algal bloom distribution map is calculated; based on the binary algal bloom distribution map, an algal bloom ROI image is extracted from the multi-band ground reflectance mosaic map.

[0006] In the above-mentioned automatic identification method of algal blooms based on multispectral information, the obtained original multispectral image sequence is subjected to standardized preprocessing and mosaic fusion to obtain a multi-band ground reflectance mosaic map, including: obtaining an original multispectral image sequence; performing radiometric calibration, atmospheric correction, orthorectification and registration on each original multispectral image in the original multispectral image sequence to obtain a preprocessed multispectral image sequence; mosaic fusion on the preprocessed multispectral image sequence to obtain the multi-band ground reflectance mosaic map, wherein each pixel position in the multi-band ground reflectance mosaic map includes the coastal blue band reflectance, the blue light band reflectance, the first green light band reflectance, the second green light band reflectance, the yellow light band reflectance, the red light band reflectance, the red edge band reflectance and the near-infrared band reflectance.

[0007] In the above-mentioned automatic identification method of algal blooms based on multispectral information, the water-vegetation absorption index of each pixel position in the multi-band ground reflectance mosaic is calculated, and based on the water-vegetation absorption index, the multi-band ground reflectance mosaic is subjected to intelligent water extraction to obtain a binary water range mask, including: based on the near-infrared band reflectance and blue light band reflectance of each pixel position in the multi-band ground reflectance mosaic, the water-vegetation absorption index of each pixel position in the multi-band ground reflectance mosaic is calculated using the following formula, the formula being: ;in, and is the reflectivity of the near-infrared band and the blue light band, Represents the water-vegetation absorption index.

[0008] Adaptive threshold segmentation is performed on the WAVI image composed of the water body-vegetation absorption index to obtain a binary water body range mask.

[0009] In the above-mentioned automatic identification method of algal blooms based on multispectral information, the WAVI image composed of the water body-vegetation absorption index is adaptively threshold segmented to obtain a binary water body range mask, including: performing histogram analysis on the WAVI image and using the Otsu algorithm to determine the optimal segmentation threshold for water body-non-water body; traversing each pixel in the WAVI image and judging whether the water body-vegetation absorption index of each pixel exceeds the optimal segmentation threshold for water body-non-water body to obtain the binary water body range mask.

[0010] In the above-mentioned automatic algal bloom identification method based on multispectral information, the band ratio index of each pixel position in the multi-band ground reflectance water body ROI map is calculated, and based on the band ratio index, the multi-band ground reflectance water body ROI map is adaptively pre-screened for potential algal bloom areas to obtain a rough mask of potential algal bloom areas, including: based on the red light band reflectance and the blue light band reflectance of each pixel position in the multi-band ground reflectance water body ROI map, the band ratio index of each pixel position is calculated using the following formula, the formula is: ;in, and is the reflectivity of red light band and blue light band, Represents the band ratio index.

[0011] Adaptive threshold segmentation is performed on the IRB image of the water body region of interest composed of the band ratio index to obtain a rough mask of the potential algal bloom area.

[0012] In the above-mentioned automatic identification method of algal blooms based on multispectral information, the multi-dimensional feature vector of each pixel position in the multi-band ground reflectance potential algal bloom area ROI map includes: coastal blue band reflectance, blue light band reflectance, first green light band reflectance, second green light band reflectance, yellow light band reflectance, red light band reflectance, red edge band reflectance, near-infrared band reflectance, band ratio index, yellowness index and reflection valley depth.

[0013] In the above-mentioned method for automatic identification of algal blooms based on multispectral information, the calculation formula of the yellowness index is: ;in, is the reflectivity of yellow light band, is the reflectivity of the second green light band and is the wavelength difference between the yellow light band and the second green light band, represents the yellowness index; the calculation formula for the reflection valley depth is: ;in, is the reflectivity of yellow light band, is the reflectivity of red light band, is the central wavelength of the red light band, is the central wavelength of the yellow light band, is the central band of the red edge band and is the red edge band reflectivity, Indicates the depth of the reflection valley.

[0014] In the above-mentioned automatic identification method of algal blooms based on multispectral information, a binary algal bloom distribution map is calculated based on the multidimensional feature vectors of each pixel position in the multi-band ground reflectance potential algal bloom area ROI map and a pre-trained machine learning classifier model, including: inputting the multidimensional feature vectors of each potential algal bloom pixel into the pre-trained machine learning classifier model to obtain an algal bloom classification map with confidence; performing decision thresholding and morphological processing on the algal bloom classification map with confidence to obtain the binary algal bloom distribution map.

[0015] In the above-mentioned automatic identification method of algal blooms based on multispectral information, the multidimensional feature vector of each potential algal bloom pixel is input into a pre-trained machine learning classifier model to obtain an algal bloom classification map with confidence, including: inputting the multidimensional feature vector of each potential algal bloom pixel into the convolutional encoder of the pre-trained machine learning classifier model to obtain a potential algal bloom classification feature map; performing global mean pooling along the channel dimension on the potential algal bloom classification feature map to obtain a potential algal bloom channel feature vector; determining a set of significant channel feature values ​​based on a comparison between the channel feature values ​​at each position in the potential algal bloom channel feature vector and a significant channel judgment threshold; extracting a set of significant feature matrices corresponding to the set of significant channel feature values ​​from the potential algal bloom classification feature map; performing spatial significance overall regression on the set of significant channel feature values ​​and the set of significant feature matrices to obtain an overall regression matrix; inputting the potential algal bloom classification feature map into the classification layer of the pre-trained machine learning classifier model to obtain an algal bloom probability matrix; and performing a re-probability correction based on a probability matrix representation on the algal bloom probability matrix based on the overall regression matrix to obtain the algal bloom classification map with confidence.

[0016] According to another aspect of the present application, a multispectral information-based algal bloom automatic identification system is provided, which is used to execute the above-mentioned multispectral information-based algal bloom automatic identification system, including: an image preprocessing and mosaicking module, which is used to perform standardized preprocessing and mosaic fusion on the acquired original multispectral image sequence to obtain a multi-band ground reflectance mosaic map; a water body intelligent extraction module, which is used to calculate the water body-vegetation absorption index of each pixel position in the multi-band ground reflectance mosaic map, and perform water body intelligent extraction on the multi-band ground reflectance mosaic map based on the water body-vegetation absorption index to obtain a binary water body range mask; a water body ROI map extraction module, which is used to extract a multi-band ground reflectance water body ROI map from the multi-band ground reflectance mosaic map based on the binary water body range mask; and a potential algal bloom area adaptive preliminary screening module, It is used to calculate the band ratio index of each pixel position in the multi-band ground reflectance water body ROI map, and perform adaptive preliminary screening of potential algal bloom areas on the multi-band ground reflectance water body ROI map based on the band ratio index to obtain a rough mask of the potential algal bloom area; a potential algal bloom area ROI map extraction module is used to extract the multi-band ground reflectance potential algal bloom area ROI map from the multi-band ground reflectance mosaic map based on the rough mask of the potential algal bloom area; an algal bloom distribution map generation module is used to calculate a binary algal bloom distribution map based on the multi-dimensional feature vector of each pixel position in the multi-band ground reflectance potential algal bloom area ROI map and a pre-trained machine learning classifier model; an algal bloom ROI image extraction module is used to extract an algal bloom ROI image from the multi-band ground reflectance mosaic map based on the binary algal bloom distribution map.

[0017] Compared with existing technologies, the multispectral information-based automatic algal bloom identification method and system provided in this application first performs standardized preprocessing and mosaic fusion on the original multispectral imagery to obtain high-quality ground reflectance data. Subsequently, by calculating the water-vegetation absorption index and performing intelligent extraction, the water body range is precisely defined, effectively eliminating interference from land and non-water areas. Furthermore, the band ratio index is calculated to adaptively screen potential algal bloom areas within the water body, thereby narrowing the scope of subsequent fine-grained classification and improving processing efficiency. Finally, multidimensional feature vectors, including spectral reflectance, band ratio index, yellowness index, and reflection valley depth, are extracted from these potential algal bloom areas and input into a pre-trained machine learning classifier model for deep learning and pattern recognition. This overcomes the limitations of traditional methods that rely solely on a single or a few empirical indices, and is able to capture the more complex and subtle spectral characteristics of algal blooms, achieving high-precision classification and discrimination. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application.

[0019] Figure 1 The figure illustrates a schematic flow chart of a method for automatically identifying algal blooms based on multispectral information according to an embodiment of the present application.

[0020] Figure 2 The figure illustrates a schematic flow chart of S1 in the method for automatic identification of algal blooms based on multispectral information according to an embodiment of the present application.

[0021] Figure 3 The figure illustrates a schematic flow chart of S2 in the method for automatic identification of algal blooms based on multispectral information according to an embodiment of the present application.

[0022] Figure 4 The figure illustrates a schematic flow chart of S4 in the method for automatic identification of algal blooms based on multispectral information according to an embodiment of the present application.

[0023] Figure 5 The figure illustrates a schematic flow chart of S6 in the method for automatic identification of algal blooms based on multispectral information according to an embodiment of the present application.

[0024] Figure 6 The figure shows a schematic block diagram of an automatic algal bloom identification system based on multispectral information according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0026] Based on this, this application provides a method for automatic identification of algal blooms based on multispectral information, such as Figure 1As shown, the automatic identification method of algal blooms based on multispectral information includes: S1, performing standardized preprocessing and mosaic fusion on the acquired original multispectral image sequence to obtain a multi-band ground reflectance mosaic map; S2, calculating the water body-vegetation absorption index of each pixel position in the multi-band ground reflectance mosaic map, and performing water body intelligent extraction on the multi-band ground reflectance mosaic map based on the water body-vegetation absorption index to obtain a binary water body range mask; S3, extracting a multi-band ground reflectance water body ROI map from the multi-band ground reflectance mosaic map based on the binary water body range mask; S4, calculating the water body-vegetation absorption index of each pixel position in the multi-band ground reflectance mosaic map. The multi-band ground reflectance water body ROI map is used to perform adaptive preliminary screening of potential algal bloom areas on the multi-band ground reflectance water body ROI map based on the band ratio index to obtain a rough mask of potential algal bloom areas; S5, based on the rough mask of potential algal bloom areas, extract the multi-band ground reflectance potential algal bloom area ROI map from the multi-band ground reflectance mosaic map; S6, based on the multi-dimensional feature vector of each pixel position in the multi-band ground reflectance potential algal bloom area ROI map and the pre-trained machine learning classifier model, calculate the binary algal bloom distribution map; S7, based on the binary algal bloom distribution map, extract the algal bloom ROI image from the multi-band ground reflectance mosaic map.

[0027] For example, in step S1, the obtained original multispectral image sequence is subjected to standardized preprocessing and mosaic fusion to obtain a multi-band ground reflectance mosaic map. It should be understood that the original remote sensing image is usually interfered with by various factors, such as the characteristics of the sensor itself, atmospheric effects, and terrain undulations. These interferences will cause image data distortion and cannot be directly used for accurate ground object identification and quantitative analysis. In order to ensure the accuracy and reliability of subsequent algal bloom identification, the original image must be strictly corrected and standardized to obtain ground reflectance data that truly reflects the spectral characteristics of the ground object, and integrate it into a complete and seamless regional coverage map, laying a solid data foundation for the subsequent algal bloom feature extraction and classification, thereby achieving effective acquisition of the growth law of the algal bloom.

[0028] In one embodiment, Figure 2As shown, the obtained original multispectral image sequence is subjected to standardized preprocessing and mosaic fusion to obtain a multi-band ground reflectance mosaic map, including: S11, obtaining an original multispectral image sequence; S12, performing radiometric calibration, atmospheric correction, orthorectification and registration on each original multispectral image in the original multispectral image sequence to obtain a preprocessed multispectral image sequence; S13, mosaic fusion on the preprocessed multispectral image sequence to obtain the multi-band ground reflectance mosaic map, wherein each pixel position in the multi-band ground reflectance mosaic map includes a coastal blue band reflectance, a blue light band reflectance, a first green light band reflectance, a second green light band reflectance, a yellow light band reflectance, a red light band reflectance, a red edge band reflectance and a near-infrared band reflectance.

[0029] Specifically, first, acquire a sequence of raw multispectral images. Based on the monitoring objectives and requirements (such as algal bloom growth patterns), select an appropriate remote sensing data source. For this application, remote sensing imagery with high spatial resolution (e.g., better than or reaching 3 meters) and a rich spectral band (e.g., four or more spectral bands) was selected. For example, a certain company's satellite constellation, with a large number of satellites in orbit and a short revisit period, can meet the needs of high-frequency monitoring. Its image data also includes multiple key bands, including coastal blue, blue, green I, green II, yellow, red, red edge, and near-infrared, providing rich spectral information for refined algal bloom identification. The acquired data consists of a series of single, potentially overlapping, raw images covering the entire study area. These images are stored as DN values ​​and accompanied by metadata files recording sensor parameters, imaging time, solar altitude, and other information.

[0030] Next, each original multispectral image in the original multispectral image sequence is subjected to radiometric calibration, atmospheric correction, orthorectification, and registration to obtain a preprocessed multispectral image sequence to eliminate radiometric distortion and geometric distortion in the image. The first step is radiometric calibration. The purpose of radiometric calibration is to convert the DN value, which has no physical meaning, into a radiometric brightness value or sensor top reflectivity with a clear physical unit. In software implementation, this process is completed by reading the calibration coefficients (such as gain and offset) provided in the image metadata file. The calculation formula is: ,in, For band The radiance value, is the original DN value of the band, and are the gain and bias coefficients corresponding to this band. Radiometric calibration can eliminate noise such as banding or zebra patterns caused by sensor response differences, making images acquired at different times and with different sensors initially comparable.

[0031] Next comes atmospheric correction. Its purpose is to eliminate the effects of electromagnetic wave scattering and absorption by atmospheric molecules and aerosols, further converting the onboard radiance values ​​or top reflectance obtained through radiometric calibration into true surface reflectance. Surface reflectance is an inherent physical property of ground objects, dependent solely on their characteristics and independent of atmospheric or lighting conditions. This is a critical step in achieving high-precision object classification and quantitative remote sensing. In software implementations, atmospheric correction typically employs physics-based modeling, such as the MODTRAN and 6S radiative transfer models. These models require inputs such as atmospheric parameters at the time of imaging (such as water vapor content, ozone concentration, and aerosol optical depth), as well as sun-target-sensor geometry. By simulating the propagation of electromagnetic waves through the atmosphere, they infer surface reflectance. For example, professional remote sensing software such as ENVI and ERDAS integrate sophisticated atmospheric correction modules such as FLAASH and QUAC, which automate this process. After completing the atmospheric correction, the pixel values ​​of the image represent the true reflectivity of the ground objects in each band, laying a solid physical foundation for the subsequent algal bloom identification based on spectral characteristics.

[0032] The next step is orthorectification and registration. The purpose of orthorectification is to eliminate pixel geometric position offsets and shape distortions caused by factors such as sensor attitude (such as sway and tilt during flight), terrain undulations, and the curvature of the earth. In software implementation, orthorectification uses an accurate digital elevation model (DEM) and the imaging model parameters of the sensor. Based on the row and column numbers of each pixel, the algorithm combines the sensor model and DEM data to calculate the precise geographic coordinates of the pixel on the real surface, and reprojects it to a standard map coordinate system (such as UTM) to generate an image that is geometrically as if it were taken vertically from directly above, namely an orthophoto. Registration refers to the precise alignment of all images in the image sequence and the images with the reference map to ensure that the pixel coordinates of the same feature in different images are exactly the same.

[0033] Finally, the preprocessed multispectral image sequence is mosaicked and fused. Due to the limited width of a single remote sensing image, multiple images are required to fully cover a large study area. The goal of mosaicking is to seamlessly stitch these standardized, preprocessed, and potentially overlapping images into a single, continuous, and complete image covering the entire study area. In software implementation, the mosaicking process first requires determining the stitching order and overlap between the images. In overlapping areas, color balancing or uniformity processing is performed to eliminate potential tonal differences. Then, an algorithm automatically generates a join line, finding a visually inconspicuous path within the overlap area as the stitching line. This path can be along linear features such as rivers and roads, or through areas of uniform texture to avoid cutting through intact features. Finally, feathering is performed on both sides of the join line to ensure a smooth transition between pixel values ​​in the two images, ultimately generating a visually and data-wise seamless mosaic—the multiband ground reflectance mosaic.

[0034] Exemplarily, in step S2, the water-vegetation absorption index of each pixel position in the multi-band ground reflectance mosaic is calculated, and based on the water-vegetation absorption index, the multi-band ground reflectance mosaic is subjected to water body intelligent extraction to obtain a binary water body range mask. It should be understood that algal blooms only occur in water bodies. Therefore, before performing algal bloom identification, accurately separating the water body area from the land, vegetation or other non-water body areas is a crucial preprocessing step. Through water body intelligent extraction, the interference of the land background can be effectively eliminated, and the subsequent algal bloom identification algorithm can be focused on the water body area, thereby greatly improving the accuracy and efficiency of identification and avoiding the misjudgment of vegetation or other land objects on land as algal blooms.

[0035] In one embodiment, Figure 3 As shown, the water-vegetation absorption index of each pixel position in the multi-band ground reflectance mosaic is calculated, and based on the water-vegetation absorption index, the multi-band ground reflectance mosaic is subjected to intelligent water extraction to obtain a binary water range mask, including: S21, based on the near-infrared band reflectance and the blue light band reflectance of each pixel position in the multi-band ground reflectance mosaic, the water-vegetation absorption index of each pixel position in the multi-band ground reflectance mosaic is calculated using the following formula, the formula being: ;in, and is the reflectivity of the near-infrared band and the blue light band, S22: performing adaptive threshold segmentation on the WAVI image composed of the water-vegetation absorption index to obtain a binary water range mask.

[0036] It should be understood that it is necessary to automatically and intelligently find an optimal segmentation threshold for the WAVI image so as to perfectly divide it into two categories: water and non-water bodies. Using a fixed, empirically set threshold (for example, setting WAVI < 0 to indicate water) is a simple but very unreliable method. Because in different remote sensing images, the specific distribution range of WAVI values ​​will change due to differences in imaging time, lighting conditions, atmospheric conditions, water clarity, and sensor characteristics. A fixed threshold may work well in some scenarios, but may lead to incorrect extraction of water body boundaries in other scenarios. Therefore, this application adopts a more advanced and robust adaptive threshold segmentation method, namely, the Otsu algorithm.

[0037] In one embodiment, adaptive threshold segmentation is performed on a WAVI image composed of the water-vegetation absorption index to obtain a binary water range mask, including: performing a histogram analysis on the WAVI image and determining an optimal water-non-water segmentation threshold using the Otsu algorithm; and traversing each pixel in the WAVI image and determining whether the water-vegetation absorption index of each pixel exceeds the optimal water-non-water segmentation threshold to obtain the binary water range mask. It should be understood that the Otsu algorithm is a classic algorithm widely used in the field of image segmentation. Its intelligence lies in its ability to automatically calculate the optimal segmentation threshold based solely on the grayscale distribution characteristics of the image itself, without requiring any prior knowledge or manual intervention. Its core concept is to traverse all possible grayscale thresholds and find the threshold that maximizes the inter-class variance between the two segmented classes (in this application's scenario, "water" and "non-water"). Maximizing the between-class variance is equivalent to minimizing the within-class variance. This means that after segmentation using this threshold, the pixel values ​​within each class are as similar as possible (i.e., the WAVI values ​​of water pixels are clustered together, and the WAVI values ​​of non-water pixels are also clustered together), while the difference between the two classes is maximized. This is completely consistent with our goal of clearly distinguishing water from non-water bodies.

[0038] For example, in step S3, a multi-band ground reflectance water ROI map is extracted from the multi-band ground reflectance mosaic map based on the binarized water range mask. It should be understood that the use of the binarized water range mask can effectively remove non-water areas such as land, vegetation, and buildings from the original multi-band ground reflectance mosaic map, thereby significantly reducing the amount of data, lowering computational complexity, and avoiding interference and misjudgment of algal bloom identification results due to non-water areas. This water-focused processing approach ensures the targeted and accurate identification of algal blooms, laying a pure data foundation for subsequent refined analysis.

[0039] Specifically, based on the binary water range mask, a multi-band ground reflectance water ROI map is extracted from the multi-band ground reflectance mosaic. The algorithm first creates a new, blank raster data file with the same size, number of bands, and georeference information as the input multi-band ground reflectance mosaic, serving as the output multi-band ground reflectance water ROI map. The algorithm then traverses the entire image space pixel by pixel. For any pixel coordinate (x, y) in the image, the algorithm first reads the pixel value of that coordinate on the binary water range mask. Next, a conditional check is performed: if the mask pixel value is 1 (indicating a water body at that location), the algorithm reads and copies the reflectance values ​​of all bands at that coordinate (x, y) from the input multi-band ground reflectance mosaic, forming a pixel vector containing the multi-band reflectance values. This pixel vector is then written to the same coordinate (x, y) in the output multi-band ground reflectance water ROI map. Conversely, if the mask pixel value is 0 (indicating that the location is not a water body), the algorithm will write a predefined invalid value to the corresponding coordinate (x, y) position in the output multi-band ground reflectance water body ROI map, for example, directly setting the pixel value of all bands to 0. The purpose of this invalid value is to clearly indicate that the pixel does not belong to the region of interest and should be ignored in subsequent statistics and analysis.

[0040] Exemplarily, in step S4, a band ratio index is calculated for each pixel position in the multi-band ground reflectance water body ROI map, and based on the band ratio index, the multi-band ground reflectance water body ROI map is adaptively pre-screened for potential algal bloom areas to obtain a rough mask of potential algal bloom areas. It should be understood that although the analysis scope has been limited to the water body, in addition to algal blooms, the water body may also contain other aquatic vegetation, suspended sediment, and other ground features, which differ spectrally from algal blooms. By introducing the band ratio index for preliminary screening, the unique spectral reflectance characteristics of algal blooms can be utilized to quickly and effectively distinguish areas in the water body where algal blooms may exist from non-algal bloom water bodies, thereby significantly narrowing the scope of subsequent fine-grained classification, improving overall recognition efficiency, and providing a purer and more targeted input dataset for subsequent more complex machine learning classifications, avoiding the input of large numbers of non-algal bloom areas into computationally intensive machine learning models, thereby optimizing the use of computing resources.

[0041] In one embodiment, Figure 4As shown, the band ratio index of each pixel position in the multi-band ground reflectance water body ROI map is calculated, and based on the band ratio index, the multi-band ground reflectance water body ROI map is adaptively pre-screened for potential algal bloom areas to obtain a rough mask of potential algal bloom areas, including: S41, based on the red light band reflectance and the blue light band reflectance of each pixel position in the multi-band ground reflectance water body ROI map, the band ratio index of each pixel position is calculated using the following formula, the formula is: ;in, and is the reflectivity of red light band and blue light band, Represents a band ratio index, S42, performing adaptive threshold segmentation on the IRB image of the water body region of interest composed of the band ratio index to obtain a rough mask of the potential algal bloom area.

[0042] Specifically, when adaptively screening potential algal bloom areas on the multi-band ground reflectance water body ROI map based on the band ratio index to obtain a rough mask of potential algal bloom areas, how to find an optimal segmentation threshold for the water body region of interest (IRB) image, thereby segmenting the image into two categories: potential algal bloom and non-algal bloom water bodies. Using a fixed empirical threshold (for example, setting IRB>1.5 to indicate algal bloom) is not advisable because the distribution range of IRB values ​​varies across time, different water areas, different algal bloom types, and even different sensors. A fixed threshold will lead to poor generalization ability and may result in large-scale missed or misjudgment in certain scenarios. Therefore, in this step, an adaptive threshold segmentation method, namely the Otsu algorithm, is also used.

[0043] Exemplarily, in step S5, based on the rough mask of the potential algal bloom area, a multi-band ground reflectance ROI map of the potential algal bloom area is extracted from the multi-band ground reflectance mosaic. It should be understood that although a rough mask of the potential algal bloom area has been generated, the mask itself is only a binary indicator map and does not contain the original multispectral reflectance data. Therefore, through this step, the multispectral data corresponding to the rough mask can be accurately cropped from the original, complete multi-band ground reflectance mosaic, thereby greatly reducing the amount of data that needs to be input into the machine learning model, significantly improving the computational efficiency of subsequent classification, and ensuring the pertinence and accuracy of the classification, avoiding unnecessary complex calculations for non-potential algal bloom areas.

[0044] Specifically, based on the coarse mask of the potential algal bloom areas, a multi-band ground reflectance ROI map of potential algal bloom areas is extracted from the multi-band ground reflectance mosaic. This means that for each pixel marked as a potential algal bloom area in the coarse mask (e.g., with a pixel value of 1), the reflectance data for all corresponding spectral bands (including the coastal blue band, blue band, first green band, second green band, yellow band, red band, red-edge band, and near-infrared band) in the original multi-band ground reflectance mosaic are fully preserved. Conversely, for pixels marked as non-potential algal bloom areas in the coarse mask (e.g., with a pixel value of 0), the corresponding pixel values ​​in the extracted multi-band ground reflectance ROI map of potential algal bloom areas are set to invalid values ​​or 0, thereby effectively eliminating data from non-potential algal bloom areas.

[0045] Exemplarily, in step S6, a binary algal bloom distribution map is calculated based on the multidimensional feature vectors of each pixel position in the multi-band ground reflectance potential algal bloom area ROI map and the pre-trained machine learning classifier model. It should be understood that although the previous processing has narrowed the analysis scope to the potential algal bloom area, these areas may still contain non-algal bloom water features, such as aquatic plants, suspended matter or clear water bodies, which may be confused with algal blooms in the spectrum. Traditional methods based on a single or a few spectral indices are often difficult to achieve high-precision and high-robustness distinction when faced with complex and changeable water environments. Therefore, the introduction of multidimensional feature vectors and advanced machine learning classifier models can fully explore the subtle differences and complex patterns of algal blooms in multiple spectral dimensions, thereby overcoming the limitations of traditional methods, achieving accurate identification and classification of algal blooms, and significantly improving the accuracy and generalization ability of recognition.

[0046] Specifically, first, it is necessary to calculate the multidimensional feature vector of each pixel position in the multi-band ground reflectance potential algal bloom area ROI map. For each pixel in the multi-band ground reflectance potential algal bloom area ROI map, its multidimensional feature vector is extracted, where the multidimensional feature vector of each pixel position in the multi-band ground reflectance potential algal bloom area ROI map includes: coastal blue band reflectance, blue light band reflectance, first green light band reflectance, second green light band reflectance, yellow light band reflectance, red light band reflectance, red edge band reflectance, near-infrared band reflectance, band ratio index, yellowness index, and reflection valley depth.

[0047] In one embodiment, the calculation formula of the yellowness index is: ;in, is the reflectivity of yellow light band, is the reflectivity of the second green light band and is the wavelength difference between the yellow light band and the second green light band, Indicates the yellowness index.

[0048] The calculation formula of the reflection valley depth is: ;in, is the reflectivity of yellow light band, is the reflectivity of red light band, is the central wavelength of the red light band, is the central wavelength of the yellow light band, is the central band of the red edge band and is the red edge band reflectivity, Indicates the depth of the reflection valley.

[0049] In one embodiment, Figure 5 As shown, based on the multidimensional feature vector of each pixel position in the multi-band ground reflectance potential algal bloom area ROI map and the pre-trained machine learning classifier model, a binary algal bloom distribution map is calculated, including: S61, inputting the multidimensional feature vector of each potential algal bloom pixel into the pre-trained machine learning classifier model to obtain an algal bloom classification map with confidence; S62, performing decision thresholding and morphological processing on the algal bloom classification map with confidence to obtain the binary algal bloom distribution map.

[0050] Specifically, a pre-trained machine learning classifier model is first used for classification prediction. This is the core of the intelligent decision-making process in this step. Pre-training means that the model is not trained instantly when processing the current image. Instead, it has been thoroughly trained and optimized using a large, high-quality, expertly annotated remote sensing imagery. This library contains a large number of pixels known to indicate algal blooms and non-algal blooms (e.g., clear water, turbid water, aquatic vegetation), along with their corresponding multi-dimensional feature vectors. By learning from this massive sample, the model has mastered the complex, nonlinear inherent laws that distinguish between algal blooms and non-algal blooms.

[0051] In the present application, convolutional neural networks are preferably used as the specific machine learning classifier model architecture. Although traditional pixel-level classification tasks often use support vector machines or random forests, CNN has a natural advantage in processing image data. It can automatically learn and extract spatial-spectral hierarchical features in the data. Even in pixel-level classification tasks, each pixel and its neighborhood can be input into CNN as a small image block. Specifically, the CNN model architecture is as follows, which includes a convolutional encoder and a classification layer, wherein the convolutional encoder applies multiple one-dimensional convolution kernels to perform sliding convolution on the multidimensional feature vector of each potential algal bloom pixel. This can capture the local combination relationship between different spectral features, such as a pattern of high red light reflectance and high yellowness index. After the convolution, a ReLU activation function is connected to introduce nonlinearity. After activation, a maximum pooling layer is performed to downsample the output of the convolution layer, reduce the data dimension, retain the most significant features, and increase the translation invariance of the model. The classification layer is a fully connected layer containing a single neuron and uses a Sigmoid activation function. The Sigmoid function can map any real-number output to the (0, 1) interval, and this output value can be directly interpreted as the probability that the pixel belongs to the algal bloom category.

[0052] During implementation, a pretrained CNN model is loaded into memory. The multidimensional feature vectors of each potential algal bloom pixel are then fed into the model one by one. After a forward propagation, the model outputs a probability value between 0 and 1 for each pixel. These probabilities are combined to form an algal bloom probability map that corresponds to the spatial extent of the potential algal bloom region of interest (ROI) map.

[0053] Here, for the multidimensional feature vectors of each potential algal bloom pixel, since not only are there correlations between the numerical values ​​of the feature vectors themselves, but there are also complex image spatial distribution correlations between the feature vectors corresponding to each pixel, it is preferred that the machine learning classifier model extracts complex correlations in width, height, and channel dimensions through a convolutional neural network. In this way, for the obtained classification feature map, considering that in the spatial dimension of the feature matrix, it has successively undergone adaptive threshold segmentation and screening guided by the eigenvalue correlation in the channel dimension, and the purpose of feature map regression is to obtain a probability matrix representation, it is expected to improve the spatial probability distribution regression of the feature map with spatial distribution screening characteristics in the channel dimension through feature map optimization.

[0054] Based on this, in another embodiment, the multidimensional feature vectors of each potential algal bloom pixel are input into a pretrained machine learning classifier model to obtain an algal bloom classification map with confidence, including: first, inputting the multidimensional feature vectors of each potential algal bloom pixel into the convolutional encoder of the pretrained machine learning classifier model to obtain a potential algal bloom classification feature map. It should be understood that although the original multidimensional feature vector contains rich spectral information, this information exists in a relatively discrete and linear manner. In order to reveal the deep, nonlinear coupling relationship between these features, as well as the contextual information of the pixel in the spatial neighborhood, the powerful feature extraction capabilities of the convolutional neural network are utilized. Through its stacked convolutional layers and nonlinear activation functions, the convolutional encoder can automatically learn and extract abstract features from low-level to high-level, such as from simple band ratio relationships to complex spectral curve morphology and spatial texture patterns. In this way, the original, low-information-density multidimensional feature vectors of each potential algal bloom pixel can be converted into a high-dimensional, highly information-concentrated potential algal bloom classification feature map.

[0055] Then, the potential algal bloom classification feature map is globally mean pooled along the channel dimension to obtain a potential algal bloom channel feature vector, and based on the comparison between the channel feature values ​​at each position in the potential algal bloom channel feature vector and the significant channel determination threshold, a set of significant channel feature values ​​is determined, that is, for the channel feature vector Each eigenvalue of ,choose The significant passage here, and The channel feature vectors are Each eigenvalue of The mean and standard deviation of the set composed of is a modulation coefficient used to avoid over-screening and under-screening. Here, the modulation coefficient, as a hyperparameter in machine learning, can be determined by an existing hyperparameter tuning method. In one embodiment, Of course, this is just an example and is not a specific limitation.

[0056] Next, a significant feature matrix corresponding to the set of significant channel eigenvalues ​​is extracted from the potential algal bloom classification feature map. The set of significant channel eigenvalues ​​and the set of significant feature matrices are subjected to spatial significance overall regression to obtain an overall regression matrix, which is expressed as: ;in, Represents each significant feature matrix, represents the eigenvalues ​​of each significant channel, Indicates that each significant feature matrix The corresponding channel eigenvalue After performing point multiplication to obtain the weighted feature matrix, all weighted feature matrices are summed up. Represents the overall regression matrix, so that due to the channel eigenvalues By encoding the overall spatial significance distribution information, the overall regression matrix can highlight the key spatial distribution with regression significance for probability regression through spatial significance overall regression, and combine the distribution screening characteristics of channel dimension and spatial dimension.

[0057] Finally, the potential algal bloom classification feature map is input into the classification layer of the pre-trained machine learning classifier model to obtain an algal bloom probability matrix, and the algal bloom probability matrix is ​​subjected to a re-probability correction based on the probability matrix representation based on the overall regression matrix to obtain the algal bloom classification map with confidence. Here, inputting the potential algal bloom classification feature map into the classification layer of the pre-trained machine learning classifier model to obtain an algal bloom probability matrix is ​​a standard process for deep learning classification, and its purpose is to generate a preliminary probability value for each pixel belonging to the algal bloom category based on all the features extracted by the encoder. The initial probability matrix is ​​then corrected based on the re-probability representation of the probability matrix through the overall regression matrix, that is, in the potential algal bloom classification feature map The algal bloom probability matrix obtained by the classifier is: In the case of , that is, the algal bloom probability matrix With the overall regression matrix The eigenvalues ​​at the same position of are used as the exponent and base respectively, and then applied Function activation is obtained by re-probabilization , and the algal bloom probability matrix Perform a dot multiplication to obtain the aforementioned algal bloom classification map with confidence. It's important to note that the initial probability matrix shrinks overall after correction, so it can be restored using maximum normalization. Through the above calculation, in areas where the overall regression matrix indicates high significance (i.e., areas where the model is most confident), the initial probabilities are significantly pushed toward 0 or 1, thereby increasing their confidence. In areas where significance is low, the probability values ​​are adjusted more modestly.

[0058] In this way, the integrated overall regression is performed in the channel dimension based on the regional regression characteristics after spatial distribution screening, thereby combining the regional significance probability prediction, strengthening the criticality of the spatial probability distribution, and ultimately improving the accuracy of the probability confidence distribution of the obtained algal bloom classification map.

[0059] Then, the algal bloom classification map with confidence is subjected to decision thresholding and morphological processing to obtain the binary algal bloom distribution map. First, decision thresholding is performed, which is a simple binary classification decision process. A probability threshold needs to be set, and usually, this threshold is set to 0.5. The algorithm will traverse each pixel in the algal bloom probability map. If its probability value is greater than 0.5, the pixel is considered to belong to the algal bloom category, and the value 1 is written to the corresponding position of the final binary algal bloom distribution map; if its probability value is less than or equal to 0.5, it is considered to belong to the non-algal bloom category, and the value 0 is written. In applications with special requirements for accuracy, this threshold can also be adjusted according to ROC curve analysis to achieve the best balance between recall rate and precision rate.

[0060] Next, morphological processing is performed. Binary images directly obtained by thresholding may contain some noise, such as isolated pixels that are misclassified (salt and pepper noise) or small holes within large algal blooms. These do not conform to the geographical pattern of algal blooms, which typically occur in patches in nature. Morphological processing aims to eliminate this noise and make the resulting image smoother and more regular. Common morphological operations include: Opening, which first erodes the image, then dilates it. Erosion "eats away" the boundaries of the object area, effectively eliminating small, isolated bright spots (single pixels or small clusters of pixels mistakenly identified as algal blooms). The subsequent dilation restores the main object area, slightly reduced by erosion, to a near-original size. The overall effect is to remove small noise points and smooth the object's outline. Closing: This operation is the opposite of Opening, performing dilation followed by erosion. Dilation fills small holes within the object and connects adjacent object areas. The subsequent erosion restores the object boundary, slightly enlarged by dilation, to fill holes within the object and connect broken areas.

[0061] Illustratively, in step S7, an algal bloom ROI image is extracted from the multi-band ground reflectance mosaic based on the binary algal bloom distribution map. It should be understood that this binary image itself is merely a logical judgment result and does not contain the original spectral reflectance information of the algal bloom area. In order to enable further quantitative analysis and visualization of the identified algal blooms, or to serve as input for other applications (such as algal bloom biomass estimation and chlorophyll concentration inversion), obtaining the original multispectral image data of these algal bloom areas is essential. Through this step, the ultimately identified algal bloom area can be accurately extracted from the original, complete multi-band ground reflectance mosaic, thereby providing an image containing the true spectral characteristics of the algal bloom, namely, the algal bloom ROI image.

[0062] Specifically, based on the binary algal bloom distribution map, an algal bloom ROI image is extracted from the multi-band ground reflectance mosaic. This means that for each pixel marked as an algal bloom area (e.g., with a pixel value of 1) in the binary algal bloom distribution map, the reflectance data for all corresponding spectral bands (including coastal blue band reflectance, blue light band reflectance, first green light band reflectance, second green light band reflectance, yellow light band reflectance, red light band reflectance, red-edge band reflectance, and near-infrared band reflectance) in the original multi-band ground reflectance mosaic are fully preserved. Conversely, for pixels marked as non-algal bloom areas (e.g., with a pixel value of 0) in the binary algal bloom distribution map, the corresponding pixel values ​​in the extracted algal bloom ROI image are set to invalid values ​​or 0, thereby accurately removing data from non-algal bloom areas.

[0063] In summary, the method for automatic algal bloom identification based on multispectral information provided in this application first performs standardized preprocessing and mosaic fusion on the original multispectral imagery to obtain high-quality ground reflectance data. Subsequently, by calculating the water-vegetation absorption index and performing intelligent extraction, the water body range is accurately defined, effectively eliminating interference from land and non-water areas. The band ratio index is further calculated to adaptively screen potential algal bloom areas within the water body, thereby narrowing the scope of subsequent fine classification and improving processing efficiency. Finally, multidimensional feature vectors, including spectral reflectance, band ratio index, yellowness index, and reflection valley depth, are extracted from these potential algal bloom areas and input into a pre-trained machine learning classifier model for deep learning and pattern recognition. This overcomes the limitations of traditional methods that rely solely on a single or a few empirical indices, and is able to capture the more complex and subtle spectral characteristics of algal blooms, achieving high-precision classification and discrimination.

[0064] The present application also provides an automatic algae bloom identification system based on multispectral information, which is used to execute the above-mentioned automatic algae bloom identification system based on multispectral information, such as Figure 6As shown, the automatic algal bloom identification system 600 based on multispectral information includes: an image preprocessing and mosaicking module 610, which is used to perform standardized preprocessing and mosaic fusion on the acquired original multispectral image sequence to obtain a multi-band ground reflectance mosaic map; a water body intelligent extraction module 620, which is used to calculate the water body-vegetation absorption index of each pixel position in the multi-band ground reflectance mosaic map, and perform water body intelligent extraction on the multi-band ground reflectance mosaic map based on the water body-vegetation absorption index to obtain a binary water body range mask; a water body ROI map extraction module 630, which is used to extract a multi-band ground reflectance water body ROI map from the multi-band ground reflectance mosaic map based on the binary water body range mask; a potential algal bloom area adaptive primary screening module 640, which is used to calculate the multi-band ground reflectance water body The multi-band ground reflectance water body ROI map is used to adaptively screen potential algal bloom areas based on the band ratio index of each pixel position in the ROI map, so as to obtain a rough mask of the potential algal bloom area; a potential algal bloom area ROI map extraction module 650 is used to extract a multi-band ground reflectance potential algal bloom area ROI map from the multi-band ground reflectance mosaic map based on the rough mask of the potential algal bloom area; an algal bloom distribution map generation module 660 is used to calculate a binary algal bloom distribution map based on the multi-dimensional feature vector of each pixel position in the multi-band ground reflectance potential algal bloom area ROI map and a pre-trained machine learning classifier model; an algal bloom ROI image extraction module 670 is used to extract an algal bloom ROI image from the multi-band ground reflectance mosaic map based on the binary algal bloom distribution map.

[0065] An embodiment of the present application also provides a computer-readable storage medium, which stores computer program code. When the computer program code runs on a computer, the computer executes the above-mentioned related method steps to implement an automatic algal bloom identification method based on multispectral information provided by the above embodiment.

[0066] An embodiment of the present application also provides a computer program product. When the computer program product is run on a computer, it enables the computer to execute the above-mentioned related steps to implement an automatic identification method of algal blooms based on multispectral information provided by the above embodiment.

[0067] Among them, the system, computer-readable storage medium or computer program product provided in the embodiments of the present application are all used to execute the corresponding methods provided above. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding methods provided above, and will not be repeated here.

[0068] It should be noted that the order of the above embodiments of the present application is only for description and does not represent the advantages or disadvantages of the embodiments.

[0069] The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous. The various embodiments in this specification are described in a progressive manner. Similar parts between the various embodiments can be referenced to each other. Each embodiment focuses on the differences from other embodiments.

Claims

1. A method for automatically identifying algal blooms based on multispectral information, characterized in that: include: The obtained original multispectral image sequence is subjected to standardized preprocessing and mosaic fusion to obtain a multi-band ground reflectance mosaic map; the water-vegetation absorption index of each pixel position in the multi-band ground reflectance mosaic map is calculated, and based on the water-vegetation absorption index, water body intelligent extraction is performed on the multi-band ground reflectance mosaic map to obtain a binary water body range mask; Based on the binarized water range mask, a multi-band ground reflectance water body ROI map is extracted from the multi-band ground reflectance mosaic map; a band ratio index of each pixel position in the multi-band ground reflectance water body ROI map is calculated, and based on the band ratio index, the multi-band ground reflectance water body ROI map is adaptively preliminarily screened for potential algal bloom areas to obtain a rough mask of potential algal bloom areas; Based on the rough mask of the potential algal bloom area, a multi-band ground reflectance potential algal bloom area ROI map is extracted from the multi-band ground reflectance mosaic map; based on the multi-dimensional feature vector of each pixel position in the multi-band ground reflectance potential algal bloom area ROI map and a pre-trained machine learning classifier model, a binary algal bloom distribution map is calculated; based on the binary algal bloom distribution map, an algal bloom ROI image is extracted from the multi-band ground reflectance mosaic map.

2. The method for automatic identification of algal blooms based on multispectral information according to claim 1, characterized in that: The method comprises the following steps: obtaining an original multispectral image sequence; performing radiometric calibration, atmospheric correction, orthorectification and registration on each original multispectral image in the original multispectral image sequence to obtain a preprocessed multispectral image sequence; and mosaicking and fusing the preprocessed multispectral image sequence to obtain the multiband ground reflectance mosaic map, wherein each pixel position in the multiband ground reflectance mosaic map includes a coastal blue band reflectance, a blue light band reflectance, a first green light band reflectance, a second green light band reflectance, a yellow light band reflectance, a red light band reflectance, a red edge band reflectance and a near infrared band reflectance.

3. The method for automatic identification of algal blooms based on multispectral information according to claim 2, characterized in that: Calculating the water-vegetation absorption index of each pixel position in the multi-band ground reflectance mosaic, and performing water intelligent extraction on the multi-band ground reflectance mosaic based on the water-vegetation absorption index to obtain a binary water range mask, including: calculating the water-vegetation absorption index of each pixel position in the multi-band ground reflectance mosaic based on the near-infrared band reflectance and the blue light band reflectance of each pixel position in the multi-band ground reflectance mosaic using the following formula, the formula being: ;in, and is the reflectivity of the near-infrared band and the blue light band, represents the water-vegetation absorption index; and performing adaptive threshold segmentation on the WAVI image composed of the water-vegetation absorption index to obtain a binary water range mask.

4. The method for automatic identification of algal blooms based on multispectral information according to claim 3, characterized in that: The method comprises the following steps: performing adaptive threshold segmentation on a WAVI image composed of the water body-vegetation absorption index to obtain a binary water body range mask, comprising: performing histogram analysis on the WAVI image and determining an optimal water body-non-water body segmentation threshold using an Otsu algorithm; and traversing each pixel in the WAVI image and determining whether the water body-vegetation absorption index of each pixel exceeds the optimal water body-non-water body segmentation threshold to obtain the binary water body range mask.

5. The method for automatic identification of algal blooms based on multispectral information according to claim 2, characterized in that: Calculating the band ratio index of each pixel position in the multi-band ground reflectance water body ROI map, and performing adaptive preliminary screening of potential algal bloom areas on the multi-band ground reflectance water body ROI map based on the band ratio index to obtain a rough mask of the potential algal bloom area, including: calculating the band ratio index of each pixel position based on the red light band reflectance and the blue light band reflectance of each pixel position in the multi-band ground reflectance water body ROI map using the following formula, wherein the formula is: ;in, and is the reflectivity of red light band and blue light band, Represents a band ratio index; adaptive threshold segmentation is performed on the IRB image of the water body region of interest composed of the band ratio index to obtain a rough mask of the potential algal bloom area.

6. The method for automatic identification of algal blooms based on multispectral information according to claim 5, characterized in that: The multi-dimensional feature vector of each pixel position in the multi-band ground reflectance potential algal bloom area ROI map includes: coastal blue band reflectance, blue light band reflectance, first green light band reflectance, second green light band reflectance, yellow light band reflectance, red light band reflectance, red edge band reflectance, near-infrared band reflectance, band ratio index, yellowness index and reflection valley depth.

7. The method for automatic identification of algal blooms based on multispectral information according to claim 6, characterized in that: The calculation formula of the yellowness index is: ;in, is the reflectivity of yellow light band, is the reflectivity of the second green light band and is the wavelength difference between the yellow light band and the second green light band, represents the yellowness index; the calculation formula for the reflection valley depth is: ;in, is the reflectivity of yellow light band, is the red light band reflectivity, is the central wavelength of the red light band, is the central wavelength of the yellow light band, is the central band of the red edge band and is the red edge band reflectivity, Indicates the depth of the reflection valley.

8. The method for automatic identification of algal blooms based on multispectral information according to claim 7, characterized in that: Based on the multidimensional feature vectors of each pixel position in the multi-band ground reflectance potential algal bloom area ROI map and a pre-trained machine learning classifier model, a binary algal bloom distribution map is calculated, including: inputting the multidimensional feature vectors of each potential algal bloom pixel into the pre-trained machine learning classifier model to obtain an algal bloom classification map with confidence; performing decision thresholding and morphological processing on the algal bloom classification map with confidence to obtain the binary algal bloom distribution map.

9. The method for automatic identification of algal blooms based on multispectral information according to claim 8, characterized in that: The multidimensional feature vector of each potential algal bloom pixel is input into a pre-trained machine learning classifier model to obtain an algal bloom classification map with confidence, including: inputting the multidimensional feature vector of each potential algal bloom pixel into the convolutional encoder of the pre-trained machine learning classifier model to obtain a potential algal bloom classification feature map; performing global mean pooling along the channel dimension on the potential algal bloom classification feature map to obtain a potential algal bloom channel feature vector; determining a set of significant channel feature values ​​based on a comparison between the channel feature values ​​at each position in the potential algal bloom channel feature vector and a significant channel determination threshold; extracting a set of significant feature matrices corresponding to the set of significant channel feature values ​​from the potential algal bloom classification feature map; performing spatial significance overall regression on the set of significant channel feature values ​​and the set of significant feature matrices to obtain an overall regression matrix; inputting the potential algal bloom classification feature map into the classification layer of the pre-trained machine learning classifier model to obtain an algal bloom probability matrix; and performing a re-probability correction based on a probability matrix representation on the algal bloom probability matrix based on the overall regression matrix to obtain the algal bloom classification map with confidence.

10. An automatic algal bloom identification system based on multispectral information, characterized in that: include: An image preprocessing and mosaicking module is used to perform standardized preprocessing and mosaic fusion on the acquired original multispectral image sequence to obtain a multi-band ground reflectance mosaic map; a water body intelligent extraction module is used to calculate the water body-vegetation absorption index of each pixel position in the multi-band ground reflectance mosaic map, and perform water body intelligent extraction on the multi-band ground reflectance mosaic map based on the water body-vegetation absorption index to obtain a binary water body range mask; a water body ROI map extraction module is used to extract a multi-band ground reflectance water body ROI map from the multi-band ground reflectance mosaic map based on the binary water body range mask; a potential algal bloom area adaptive preliminary screening module is used to calculate the band ratio index of each pixel position in the multi-band ground reflectance water body ROI map, and perform potential algal bloom area adaptive preliminary screening on the multi-band ground reflectance water body ROI map based on the band ratio index to obtain a rough mask of potential algal bloom areas; A potential algal bloom area ROI map extraction module is used to extract a multi-band ground reflectance potential algal bloom area ROI map from the multi-band ground reflectance mosaic map based on a rough mask of the potential algal bloom area; an algal bloom distribution map generation module is used to calculate a binary algal bloom distribution map based on the multi-dimensional feature vector of each pixel position in the multi-band ground reflectance potential algal bloom area ROI map and a pre-trained machine learning classifier model; an algal bloom ROI image extraction module is used to extract an algal bloom ROI image from the multi-band ground reflectance mosaic map based on the binary algal bloom distribution map.

Images