A hyperspectral imaging detection system for marine plankton

Abstract

This invention discloses a hyperspectral imaging detection system for marine plankton, belonging to the field of spectral detection technology, to solve the problem of poor spectral classification and recognition in aquatic environments. The invention achieves precise alignment between images and disturbances by simultaneously acquiring hyperspectral images and environmental disturbance parameters, and constructing a disturbance mapping sequence. A disturbance contribution weight matrix is ​​constructed based on the dominant band, and pixel-level spectral stripping is performed to extract the plankton purification spectrum, improving spectral purity and recognition accuracy. Furthermore, by combining derivative spectrum changes and image texture features, abnormal spectral drift regions are accurately identified. The coupling relationship between drift regions and background disturbances is further established to generate a dynamic interference factor map, dynamically adjusting the recognition model parameters, thus improving the system's recognition stability and classification accuracy in complex disturbance environments, making it suitable for multi-scenario marine ecological monitoring.

Description

Technical Field

[0001] This invention relates to the field of spectral detection technology, and more specifically, to a hyperspectral imaging detection system for marine plankton. Background Technology

[0002] Currently, marine plankton, as key primary producers in marine ecosystems, significantly reflect dynamic changes in the aquatic environment through their abundance, population distribution, and physiological state. In practical applications, hyperspectral imaging technology, with its continuous spectral acquisition and spatial information representation capabilities, has become an important means of non-contact plankton identification. Existing technologies utilize buoys, unmanned surface vessels, or remote sensing platforms equipped with hyperspectral sensors to perform pushbroom imaging of the water surface, extracting biological pigments and structural features in the visible-near-infrared band. This information, combined with image data, enables automatic identification and classification of plankton. This method offers advantages such as high real-time performance, good scene adaptability, and non-destructive monitoring, making it suitable for scenarios such as online marine ecological assessment, red tide early warning, and trophic level monitoring.

[0003] Existing technologies have limitations: Hyperspectral detection of plankton in typical nearshore or eutrophic waters still faces significant technical bottlenecks. Particularly against the backdrop of significant fluctuations in SPM (suspended particulate matter) and CDOM (dissolved organic matter) concentrations, the strong band selectivity of the scattering and absorption behavior of these non-target factors means that their spectral interference components highly overlap with the intrinsic reflectance characteristics of plankton. This makes it difficult for traditional spectral correction methods (such as global atmospheric correction or water body radiation normalization) to effectively remove interference signals. This problem manifests in spectral data as spectral slope distortion, local peak-valley shifts, and amplified normalization errors, further affecting subsequent classification based on spectral similarity or feature templates. Consequently, classification accuracy does not improve with increasing sample size; instead, it introduces the risk of overfitting, ultimately leading to decreased reliability and reduced versatility of plankton identification systems in complex aquatic environments. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, the following solution is proposed to address the problem of poor spectral classification and recognition in aquatic environments as described in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A hyperspectral imaging detection system for marine plankton includes a marine water data acquisition module, a spectral mapping module, a purification spectrum construction module, a spectral change analysis module, a disturbance generation module, and an adjustment recognition module, with each module connected by a signal.

[0007] The marine water body data acquisition module is used to acquire hyperspectral image sequences and environmental disturbance parameters of the target marine water body, perform time-series synchronous calibration, and generate disturbance mapping sequences.

[0008] The spectral mapping module is used to extract the spectral perturbation morphological features of the dominant bands of suspended particulate matter and dissolved organic matter in the perturbation mapping sequence, construct the perturbation contribution weight matrix and map it to the image pixel space;

[0009] The purification spectrum construction module is used to perform pixel-level spectral stripping based on the perturbation contribution weight matrix, extract the purification spectral curve of the planktonic target region, and construct the purification spectral feature set.

[0010] The spectral change analysis module is used to analyze the trend of derivative spectrum changes in the purified spectral feature set, and, in combination with local image texture intensity, locate regions of abnormal spectral drift.

[0011] The disturbance generation module is used to establish the spatial coupling relationship between the spectral anomaly drift region and the background disturbance source of the water body, and generate a dynamic disturbance factor spectrum;

[0012] The adjustment and identification module is used to adjust the weight distribution and target classification threshold of the spectral identification model according to the steady-state fit between the dynamic interference factor spectrum and the purification spectral feature set, so as to achieve adaptive planktonic identification under different interference conditions.

[0013] In a preferred embodiment, the marine water body data acquisition module includes:

[0014] An integrated observation platform equipped with hyperspectral imaging and environmental disturbance parameter acquisition was deployed in the target sea area;

[0015] Hyperspectral imaging acquires a sequence of surface water images at fixed time intervals and records the timestamp and geographic coordinates of each frame.

[0016] The environmental disturbance parameter acquisition system collects real-time data on suspended particulate matter concentration, dissolved organic matter concentration, incident light intensity, and wave disturbance values, generating a disturbance parameter stream.

[0017] The time synchronization controller aligns the image sequence with the disturbance parameter stream one by one using a unified time reference to construct a time synchronization structure;

[0018] Each image pixel is combined with the perturbation parameters at the corresponding time point to form a perturbation mapping sequence.

[0019] In a preferred embodiment, the spectral mapping module includes:

[0020] In the perturbation mapping sequence, spectral response analysis is performed on the perturbation parameters bound to each image pixel to extract the dominant bands of suspended particulate matter and dissolved organic matter;

[0021] Based on the dominant band, the spectral shift value, band slope change value and reflectance fluctuation rate are calculated to construct the perturbation morphology feature vector;

[0022] The perturbation morphological feature vector is normalized to generate the perturbation contribution ratio.

[0023] The perturbation contribution ratio is mapped to the image cell index to construct a perturbation contribution weight matrix, which represents the distribution weight of each perturbation factor in the image space.

[0024] In a preferred embodiment, the purification spectrum construction module includes:

[0025] For each image pixel, a perturbation spectral estimation model is constructed using the perturbation contribution weight matrix, and the perturbation components are fitted to the original reflectance spectrum.

[0026] The estimated perturbation spectrum is stripped from the original reflection spectrum band by band to output the purified spectrum;

[0027] Intensity normalization and band alignment are performed on the purified spectrum to generate a standardized spectral curve;

[0028] The standardized spectral curves of all pixels are used to form a purified spectral feature set.

[0029] In a preferred embodiment, the spectral change analysis module includes:

[0030] First-order derivative analysis and sliding curvature extraction are performed on each spectral curve in the purification spectral feature set to identify the peak shift position and the magnitude of change.

[0031] By comparing the band positions in the curvature abrupt change range with the original peak positions, the drift bands and directions are extracted.

[0032] Based on the curvature abrupt change and the texture intensity value of the corresponding extracted pixel, a consistency evaluation index between spectral derivative and texture response is constructed.

[0033] The pixel region is analyzed based on the consistency evaluation index, and the spectral anomaly drift region of the pixel region is marked.

[0034] In a preferred embodiment, a consistency evaluation index between the spectral derivative and the texture response is constructed based on the curvature abrupt change and the image texture intensity value of the corresponding extracted pixel, including:

[0035] At the image pixel corresponding to the spectral drift band, the texture intensity value calculated by the gray-level co-occurrence matrix is ​​extracted and a mapping relationship is established with the curvature abrupt value.

[0036] The consistency evaluation index is obtained by dividing the spectral curvature abrupt change value of the band corresponding to each image pixel by the image texture intensity value at the same location.

[0037] In a preferred embodiment, the disturbance generation module includes:

[0038] Establish a pixel-level registration relationship between the spatial coordinate map of the spectral anomaly drift region and the spatial distribution map of each perturbation factor in the perturbation mapping sequence;

[0039] Calculate the difference in perturbation intensity and the relative rate of change between each drift pixel and the corresponding perturbation factor to form the perturbation response gradient value;

[0040] The correlation between the perturbation response gradient value and the trend of the derivative spectrum is analyzed to calculate the perturbation coupling strength.

[0041] The perturbation coupling strength is written into the image cell index to construct a dynamic perturbation factor map.

[0042] In a preferred embodiment, the perturbation coupling strength is obtained by calculating the Pearson correlation coefficient between the perturbation response gradient value and the derivative spectrum change trend of each image pixel in the spectral anomaly drift region. The perturbation response gradient value is the change in the intensity of the perturbation factor per unit time, and the derivative spectrum change trend is the first derivative sequence of the reflectance change in continuous bands.

[0043] In a preferred embodiment, the perturbation coupling strength is written under the image cell index to construct a dynamic perturbation factor map, including:

[0044] Establish the registration relationship between the image pixel index of the spectral anomaly drift region and the spatial distribution map of each perturbation factor in the perturbation mapping sequence;

[0045] The perturbation response gradient value is obtained by multiplying the perturbation intensity difference of each drift image pixel under the corresponding perturbation factor with the perturbation change rate.

[0046] The perturbation coupling strength is obtained by calculating the correlation between the perturbation response gradient value and the change value of the first derivative in the spectral derivative curve within a sliding window.

[0047] The perturbation coupling strength is written into the image cell index to generate a dynamic perturbation factor map, which is used to characterize the intensity of the perturbation source's influence on each cell.

[0048] In a preferred embodiment, the adjustment recognition module includes:

[0049] Distribution patterns and frequencies of high-disturbance-intensity regions in the statistical dynamic disturbance factor spectrum;

[0050] Adjust the weight coefficients of the corresponding image regions in the planktonic identification model according to the interference intensity distribution;

[0051] After performing model prediction on the purification spectral feature set, the confidence bias between the predicted label and the historical stable label is calculated;

[0052] Based on the combined results of confidence bias and spatial interference factor, the classification threshold is dynamically updated to enhance target category recognition under perturbation conditions.

[0053] The technical effects and advantages of the hyperspectral imaging detection system for marine plankton of the present invention are as follows:

[0054] This invention, by simultaneously acquiring hyperspectral image sequences and environmental disturbance parameters such as suspended particulate matter, dissolved organic matter, incident light intensity, and wave disturbance in the target sea area, and calibrating them under a unified time reference, can accurately establish the mapping relationship between images and environmental disturbances, effectively ensuring data temporal consistency. By extracting the dominant band disturbance morphology of suspended particulate matter and dissolved organic matter, constructing a disturbance contribution weight matrix and mapping it to the image space, quantitative analysis of different interference factors is achieved. After performing disturbance stripping at the image pixel scale, a purified spectral feature curve is generated, which can significantly improve the purity of the spectral data on which subsequent plankton identification is based and reduce the identification error rate.

[0055] By further combining the trend of derivative spectrum changes with image texture intensity to locate spectral anomaly drift regions, subtle spectral shifts can be effectively identified, enhancing the ability to perceive disturbance-induced anomalies. By establishing the coupling relationship between spectral anomaly regions and disturbance sources, a dynamic interference factor map is generated, supporting the modeling and evaluation of the interference level in the current identification scenario. Finally, by combining the distribution of interference factors and prediction bias, the weights and classification thresholds of the identification model are dynamically adjusted, enabling stable identification and enhanced classification of plankton species under different disturbance environments. This makes it more suitable for real-time monitoring and ecological surveys of various marine plankton species. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the structure of a hyperspectral imaging detection system for marine plankton according to the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] In order to achieve the above objectives, Figure 1A schematic diagram of the structure of a marine planktonic hyperspectral imaging detection system of the present invention is provided, which specifically includes a marine water data acquisition module, a spectral mapping module, a purification spectrum construction module, a spectral change analysis module, a disturbance generation module, and an adjustment recognition module. The modules are connected by signals.

[0059] The marine water body data acquisition module is used to acquire hyperspectral image sequences and environmental disturbance parameters of the target marine water body, perform time-series synchronous calibration, and generate disturbance mapping sequences.

[0060] The spectral mapping module is used to extract the spectral perturbation morphological features of the dominant bands of suspended particulate matter and dissolved organic matter in the perturbation mapping sequence, construct the perturbation contribution weight matrix and map it to the image pixel space;

[0061] The purification spectrum construction module is used to perform pixel-level spectral stripping based on the perturbation contribution weight matrix, extract the purification spectral curve of the planktonic target region, and construct the purification spectral feature set.

[0062] The spectral change analysis module is used to analyze the trend of derivative spectrum changes in the purified spectral feature set, and, in combination with local image texture intensity, locate regions of abnormal spectral drift.

[0063] The disturbance generation module is used to establish the spatial coupling relationship between the spectral anomaly drift region and the background disturbance source of the water body, and generate a dynamic disturbance factor spectrum;

[0064] The adjustment and identification module is used to adjust the weight distribution and target classification threshold of the spectral identification model according to the steady-state fit between the dynamic interference factor spectrum and the purification spectral feature set, so as to achieve adaptive planktonic identification under different interference conditions.

[0065] The marine water body data acquisition module specifically includes:

[0066] An integrated observation platform is deployed in the target sea area to be monitored. The platform is fixed on a floating buoy or a low-speed unmanned surface vessel and is equipped with a hyperspectral imaging module and an environmental disturbance parameter acquisition module. The platform site selection should take into account typical plankton distribution zones, such as nutrient-rich areas or areas with a history of high red tide occurrence, and should ensure that the observation platform has stable hovering and time synchronization capabilities.

[0067] The hyperspectral imaging module performs spectral imaging of the sea surface at fixed time intervals, collecting wavelengths covering the visible to near-infrared range, for example, from 400 nm to 1000 nm, with an interval of 5 nm, totaling 121 wavelengths. The system continuously samples images according to a preset time frequency (e.g., one frame every 30 seconds). Each frame automatically records the acquisition time and geographic coordinates, with time accuracy at the second level and spatial positioning error less than 1 meter. The spatial resolution of the image can be set to each pixel corresponding to a 0.5 m × 0.5 m water surface area, with an image size of 512 × 512 pixels.

[0068] Simultaneously, the environmental disturbance parameter acquisition module was activated in tandem with hyperspectral imaging, collecting on-site water disturbance information at equal intervals. The acquired disturbance parameters included: suspended particulate matter concentration, dissolved organic matter concentration, vertical incident light intensity at the water surface, and surface wave disturbance value. Specifically, the suspended particulate matter concentration was obtained using a laser turbidimeter, measured in milligrams per liter; the dissolved organic matter concentration was obtained by comparing fluorescence detection with a standard response curve, measured in parts per billion; the incident light intensity was acquired by a photoelectric detection element, measured in watts per square meter; and the wave disturbance value was measured by an accelerometer combined with a water level gauge, measured in centimeters. The sampling period for these disturbance parameters was completely consistent with the hyperspectral image acquisition period and had a unified timestamp label.

[0069] To achieve synchronous matching of image data and disturbance parameters, the system is equipped with a unified time controller as a time reference. The image sequence and the disturbance parameter stream are registered one-to-one on the time axis to form a synchronization structure. Each frame of the image is bound to a set of disturbance parameters acquired at the corresponding time and labeled with the corresponding frame number, acquisition time and spatial location.

[0070] The perturbation parameter data of each frame image is combined with the perturbation parameter data of its respective frame to form an image perturbation parameter mapping relationship. For each pixel, its corresponding reflectance spectrum data is combined with the perturbation parameter values ​​(i.e., suspended particulate matter concentration, dissolved organic matter concentration, incident light intensity, and wave perturbation) at the same time point to form a perturbation mapping unit. The perturbation mapping units corresponding to all pixels in all frames of images are arranged in chronological order to form a perturbation mapping sequence. This perturbation mapping sequence is a triple data structure containing spatial location, spectral features, and synchronization perturbation factor, which can be used for subsequent perturbation stripping, spectral line purification, and target biometric identification steps.

[0071] The spectral mapping module includes:

[0072] Based on the constructed perturbation mapping sequence, perturbation spectral response analysis is performed on each image pixel to extract the influence features of suspended particulate matter and dissolved organic matter on the hyperspectral reflectance curve, and a perturbation contribution weight matrix is ​​generated for subsequent spectral stripping process.

[0073] In the perturbation mapping sequence, the system uses image pixels as index units, reading the perturbation parameter values ​​and corresponding spectral reflectance curves associated with each pixel one by one. For each perturbation factor (i.e., suspended particulate matter concentration and dissolved organic matter concentration), the system performs band-level perturbation response analysis. This process uses a fixed perturbation scanning range, such as 400 nm to 800 nm, to compare the changes in pixel reflectance spectral morphology under different time points when the perturbation parameters rise or fall, thereby identifying the band range most sensitive to the perturbation response. This band is defined as the dominant band for the corresponding perturbation factor. For example, in a set of samples, when the dissolved organic matter concentration increases from 100 ppb to 250 ppb, the pixel reflectance decreases most significantly in the 620–700 nm range; the system therefore identifies this range as the dominant band for dissolved organic matter.

[0074] Numerical extraction of spectral morphology changes in the dominant band is performed. Based on the reflectance curve of each pixel, three key perturbation characteristic indicators are calculated within the dominant band: spectral shift value, representing the degree of shift in the wavelength position of the main peak or concave point under perturbation, in nanometers; band slope change value, representing the rate of reflectance change within a unit wavelength range, used to reflect the steepening of the curve caused by perturbation; and reflectance fluctuation rate, defined as the percentage of the difference between the maximum and minimum reflectance in the dominant band to the average reflectance of the entire band, used to characterize the degree of spectral amplitude change caused by perturbation. For example, in a certain pixel, a 15-nanometer shift of the main peak, a 1.2-fold increase in slope, and a reflectance fluctuation rate of 32% were detected in the dominant band of dissolved organic matter. These three values ​​together constitute the perturbation morphology feature vector.

[0075] The perturbation morphology feature vectors are normalized to ensure that all dimensions are compared under a unified dimension. Normalization uses the maximum and minimum values ​​of the entire image area within the current monitoring period as calibration benchmarks, employing a linear mapping method to normalize all indicators to the 0–1 range. For example, when the reflectance volatility ranges from 10% to 40% across the entire image area, a pixel with a volatility of 25% has a normalization result of 0.5. The normalized three-dimensional perturbation morphology features are further integrated into a perturbation contribution ratio, which characterizes the spectral influence of a specific perturbation factor in that pixel; a larger value indicates a stronger perturbation influence.

[0076] The perturbation contribution ratios are mapped to the image space to construct a perturbation contribution weight matrix. This matrix uses the row and column coordinates of image pixels as index units, recording the perturbation contribution ratios of suspended particulate matter and dissolved organic matter pixel by pixel, and storing them in the system storage module in matrix form. Each position in the matrix represents an image pixel, and the value at that position is the spectral influence weight of the perturbation factor at that pixel. For example, at pixel 320 row and 200 column, the perturbation contribution ratio of suspended particulate matter is 0.72, and that of dissolved organic matter is 0.46. In this case, during the subsequent spectral stripping process, the spectral influence caused by suspended particulate matter will be preferentially considered for fitting and subtracting.

[0077] The marine water data acquisition module achieves precise temporal matching between images and disturbance factors by jointly acquiring hyperspectral images and environmental disturbance parameters and using a unified time reference to complete data alignment. This ensures the correlation between spectral features and disturbance behavior in subsequent analysis and enhances the system's ability to interpret marine information in dynamic environments.

[0078] The method for extracting the dominant band is as follows:

[0079] In the perturbation mapping sequence, each image pixel is bound to a set of perturbation parameters (e.g., suspended particulate matter concentration or dissolved organic matter concentration) and a corresponding hyperspectral reflectance curve. By traversing the coupling relationship between changes in perturbation parameters and spectral response, the system determines which spectral bands a particular perturbation factor significantly affects. Specifically:

[0080] A sampling sequence for a perturbation factor is selected, such as the range of dissolved organic matter concentration changes recorded during the sampling period. Representative samples of varying perturbation intensities (e.g., low, medium, and high concentration groups) are extracted from all image pixels. For each group of samples, its average reflectance curve is calculated. In each spectral band (e.g., 400–1000 nm), the reflectance difference between samples with different perturbation intensities is calculated, and the continuous band with the largest reflectance difference is identified as the dominant band for that perturbation factor. For example, if the average reflectance change exceeds 5% when dissolved organic matter changes from 100 ppb to 250 ppb in the 620–700 nm band, while it is less than 2% in other bands, then the system will consider 620–700 nm as the dominant band for dissolved organic matter.

[0081] Spectral shift describes the magnitude of the shift in the wavelength position of the main peak (or valley) within the dominant band before and after a change in perturbation conditions. It reflects the impact of perturbation on the main spectral features and is calculated as follows: Under two states of low and high perturbation intensity, the reflectance curves within the dominant band of the pixel are extracted. The position of the main peak or valley value (i.e., the wavelength where the maximum or minimum reflectance value is located) is found in each curve, and the difference in wavelength at that position between the two states is calculated. For example, if the valley appears at 660 nm under low perturbation and at 675 nm under high perturbation, the spectral shift value is 15 nm.

[0082] The band slope change value reflects the degree of change in the slope of the spectral curve within the dominant band. It is used to measure whether the disturbance makes the curve steeper or flatter. The calculation method is as follows: Select equally spaced pairs of adjacent wavelength points in the dominant band; calculate the reflectance slope of each pair of adjacent wavelength points (i.e., the reflectance difference between the two points divided by the wavelength interval); for the two states before and after the disturbance, calculate the mean of the slope respectively, and use the difference between the two as the band slope change value.

[0083] Reflectance volatility is used to characterize the intensity of the fluctuation of the reflectance curve within the dominant band, i.e. the amplitude of the spectral line. It is calculated as follows: record the maximum and minimum reflectance values ​​within the dominant band, calculate the difference between the two to obtain the reflectance amplitude, and divide the amplitude by the average reflectance value of the band to obtain the reflectance volatility, expressed as a percentage.

[0084] The perturbation morphology feature vector is constructed as follows: the three feature values ​​calculated for each image pixel's corresponding perturbation factor in the dominant band—spectral shift value, band slope change value, and reflectance fluctuation rate—are combined into a three-dimensional vector in a fixed order; the values ​​of each dimension must maintain unit consistency and stable arrangement order; for example, the three perturbation features of a certain pixel in the dissolved organic matter dominant band are 12nm, +0.003 / nm, and 27.5%, respectively.

[0085] The spectral mapping module extracts the spectral response features of the dominant perturbation factors and constructs a perturbation contribution weight matrix to achieve pixel-level perturbation weight expression. This enables the system to accurately separate the interference of environmental disturbances on the observed spectrum, providing a high-precision basis for perturbation weights for the subsequent purification process and enhancing the system's ability to distinguish complex water backgrounds.

[0086] The purification spectrum building module includes:

[0087] For each image pixel, the perturbation contribution weight matrix constructed in the previous stage is first invoked. This matrix represents the proportion of perturbation contribution of suspended particulate matter (SPM) concentration and dissolved organic matter (CDOM) concentration to the change in spectral reflectance in each dominant band. Specifically, for a specific band under each pixel index, such as 620–700 nm, the perturbation contribution weight matrix stores the perturbation weight values ​​of the two perturbation factors in that band range (e.g., SPM is 0.65, CDOM is 0.25, and the remainder is plankton and noise interference). These values ​​are obtained through normalization, and their numerical range is limited to between 0 and 1, with a sum not exceeding 1.

[0088] Based on this perturbation weight information, the system establishes a perturbation spectral estimation model on the original hyperspectral reflectance curve of the pixel. The perturbation spectral estimation model adopts a data-driven linear approximation method: the perturbation factor and the perturbation morphology feature vector of its dominant band are weighted according to their respective weights, and the perturbation spectrum caused by the perturbation factor is fitted. For example, for the 620–700 nm band, if the SPM contribution reflectance morphology derived from the perturbation morphology feature on the pixel is [0.05, 0.06, 0.08], and the CDOM is [0.03, 0.04, 0.05], then its linear perturbation spectral estimation is 0.65 × SPM morphology + 0.25 × CDOM morphology, and the fitted perturbation value on the corresponding band is [0.0415, 0.051, 0.067].

[0089] Within the same spectral band, the system strips the estimated perturbation reflectance value from the original reflectance spectrum band by band. That is, the purified spectral value at each wavelength is the original reflectance value minus the corresponding estimated perturbation contribution. For example, if the original reflectance of a certain band is 0.32 and the estimated perturbation spectral value is 0.067, then the purified reflectance for that band is 0.253. This operation ensures that the main spectral interference components of SPM and CDOM are removed from each pixel, thus leaving only the main contribution of plankton to the spectrum.

[0090] To ensure the comparability of purified spectra across different pixels, intensity normalization and band alignment were performed on all stripped purified spectra. Intensity normalization involves scaling each purified spectral curve proportionally to its maximum reflectance, ensuring all curves have a uniform maximum value of 1, thus eliminating intensity shifts caused by variations in imaging brightness. Band alignment involves interpolating all purified spectra according to standard wavelength nodes (e.g., every 5nm, from 400nm to 1000nm) to ensure consistent structure and facilitate subsequent statistical analysis. The standardized purified spectral curves from all image pixels were collected to form a purified spectral feature set. This feature set is a two-dimensional matrix structure, with rows representing image pixels and columns representing purified reflectance values ​​at standard wavelength points. This feature set can serve as the input data foundation for subsequent tasks such as plankton spectral clustering analysis, supervised classification and identification, and population lineage inference.

[0091] The purification spectrum construction module uses pixel-level perturbation stripping to remove perturbation components while ensuring the integrity of the spectral structure. This module outputs a purified, standardized spectral curve, thereby forming a basic set of spectral features of plankton and improving the specificity and accuracy of the identification algorithm for the target organism.

[0092] The spectral variation analysis module includes:

[0093] After the purification spectral feature set is constructed, the system performs derivative spectrum analysis on each standardized purification spectral curve. The derivative spectrum refers to the first derivative sequence of the reflectance of the spectral curve as a function of wavelength, which reflects the rate of change of spectral morphology.

[0094] A sliding window method is used to calculate the slope of the spectral curve between adjacent wavelengths, and the average slope value of the region is recorded at the center wavelength point of each window, thus obtaining a smooth first derivative curve. For example, for a spectral curve with a wavelength range of 400–1000 nm and an interval of 5 nm, a window size of 3 wavelength points is used. The slope at the center point of 500 nm is the average slope of the wavelengths before and after it. The degree of fluctuation in the derivative spectrum directly indicates the abrupt change in reflectance and is the basis for identifying spectral peak shifts.

[0095] The sliding curvature of each band point is calculated based on the first derivative curve. Sliding curvature describes the degree of bending in the derivative spectrum, reflecting directional abrupt changes in the spectral line within a local region. This is achieved by applying a sliding difference operation to the derivative spectrum curve again and statistically analyzing the magnitude of the change in slope between adjacent derivatives. If the curvature value of a band point exceeds a set threshold (e.g., twice the local average), it is considered that there is a possibility of spectral peak position drift at that point, and the band position and direction of this abrupt change (shift towards shorter wavelengths or longer wavelengths) are recorded.

[0096] The marked peak shift locations are compared line by line with the original peak locations of the purified spectrum. The specific peak shift location and its direction of change are extracted from each spectral curve by comparing the difference between the band index of the curvature shift point and the index of the original peak. If a peak was originally located at 450 nm in the 400–500 nm range, but the curvature shift point is at 462 nm, it indicates that the peak has shifted towards longer wavelengths, with a shift amount of 12 nm. This shift amount will be used to further determine whether it is an abnormal change.

[0097] Simultaneously, local image texture intensity values ​​are extracted from the corresponding image pixels. These texture intensity values ​​characterize the image spatial structure stability of the region where the pixel is located, and are implemented using the contrast index or local variance of the gray-level co-occurrence matrix. Specifically, a high texture intensity in a region indicates drastic changes in edges, texture, or structure; conversely, low texture intensity indicates stable image content and weak structural changes in that region.

[0098] The system further constructs a consistency evaluation index between spectral derivatives and image texture response. This index measures whether abrupt changes in spectral peak curvature are accompanied by changes in image texture. If a pixel exhibits a drastic curvature change in the derivative spectrum (i.e., a significant spectral peak shift), but the corresponding region has a very low image texture intensity (i.e., the image region itself is relatively uniform and stable), then this shift is more likely an aberrant shift caused by spectral perturbation than a genuine change in biological or structural boundaries. Therefore, this consistency index is evaluated using the inverse relationship between the curvature shift amplitude and texture intensity. Only pixels with large curvature changes and small texture changes that meet certain threshold conditions (e.g., curvature shift value greater than 0.15 and texture intensity less than 0.05) are marked as spectral aberration regions. For example, in a certain image pixel, the derivative spectrum curvature shift value is 0.23, much higher than the local average of 0.08, while its texture intensity is 0.03, lower than the overall image average texture of 0.07. This indicates that there is a strong spectral change at this location without texture support, and it is judged as an aberrant shift.

[0099] The specific method for constructing the consistency evaluation index is as follows:

[0100] For each purified spectral curve in the purified spectral feature set, its first derivative curve is extracted, and the sliding curvature value of each band point in the derivative curve is calculated to characterize the intensity of local abrupt changes in the spectral lines near that band. For each image pixel, the corresponding image texture intensity value is extracted. This texture intensity value is obtained by extracting the local gray-level variance based on a fixed window in the hyperspectral image gray-level projection map or by calculating the contrast index based on the gray-level co-occurrence matrix, to measure the degree of structural change in the pixel's neighborhood. The sliding curvature value and texture intensity value are jointly calculated to construct a consistency evaluation factor. The consistency evaluation factor is defined as the proportional relationship or correlation strength between the sliding curvature value and the texture intensity value, used to determine whether spectral abrupt changes are supported by the image's spatial structure. When the consistency evaluation factor is higher than a set threshold, and the image texture intensity is in a low range (e.g., below the 75th percentile of the mean of the entire sample image), the spectral abrupt change is considered to lack texture response support and is highly likely to be an abnormal drift. This consistency evaluation factor serves as an important screening indicator for subsequent identification of spectral abnormal drift regions, used to eliminate non-structural spectral abrupt changes caused by environmental disturbances.

[0101] The process of identifying spectral anomaly shift regions is as follows:

[0102] For each standardized spectral curve in the purified spectral feature set, the spectral derivative values ​​between continuous bands are calculated. Within a fixed-length sliding window, the spectral curvature of each band is extracted to form a spectral derivative variation curve. Curvature abrupt change intervals are identified within this curve, and their positions are compared with the peak positions in the original spectral curve to extract spectral drift bands and drift directions. At the image pixel corresponding to the spectral drift band, the texture intensity value calculated from the gray-level co-occurrence matrix is ​​extracted and a one-to-one mapping relationship is established with the curvature abrupt change value. Based on the ratio between the curvature abrupt change value and the texture intensity value, a consistency evaluation index between the spectral derivative and texture response is constructed. When the consistency evaluation index is higher than a preset threshold and the texture intensity value is lower than a set reference value, the corresponding pixel region is marked as a spectral abnormal drift region. Specifically, the ratio between the curvature abrupt change value and the texture intensity value is calculated as the consistency evaluation index. This involves dividing the spectral curvature abrupt change value of the band corresponding to each image pixel by the image texture intensity value at the same location to obtain the consistency index value for that pixel. This consistency index reflects the degree of response difference when there are drastic spectral fluctuations but no significant changes in texture, and is used to locate potential spectral drift anomaly regions. For example, if a pixel has a curvature abrupt change value of 0.18 and a texture intensity of 0.06 in a specific band, then the consistency index value is 3.0; when the consistency index value is greater than 2.5 and the texture intensity value is less than 0.08, it is classified as a drift anomaly.

[0103] The spectral change analysis module combines the trend of derivative spectrum changes with image texture intensity information to achieve highly sensitive localization of subtle spectral drift phenomena. It is particularly suitable for identifying regions with complex interference, stable textures, but abnormal spectral responses, thereby improving the system's early perception capability of latent interference behavior.

[0104] The disturbance generation module includes:

[0105] The image spatial coordinates of the identified spectral anomaly drift regions are registered pixel by pixel with the spatial distribution maps of each perturbation factor in the perturbation mapping sequence. The registration process is based on the image coordinate index alignment method to ensure that each spectral drift pixel is associated with the perturbation factor maps such as the suspended particulate matter concentration distribution map, dissolved organic matter concentration map, and fluctuation perturbation map at the same location.

[0106] After registration, the perturbation intensity value of each anomalous drift pixel in the perturbation mapping sequence is extracted sequentially, and the difference between this value and the perturbation intensity in the neighboring region is calculated to obtain the perturbation intensity difference. Simultaneously, the magnitude of the perturbation parameter change on the time axis is analyzed and divided by the change duration to obtain the relative rate of change of the perturbation. The perturbation intensity difference and the relative rate of change are used as the perturbation response gradient value to characterize the influence characteristics of the perturbation source in the anomalous region.

[0107] Subsequently, a correlation analysis was performed on the perturbation response gradient value of each pixel and the curvature abrupt change value in the first derivative spectrum of its purified spectral characteristic curve. The correlation coefficient was used to measure the degree of response coupling between the two, and the perturbation coupling strength value of the pixel was output. For perturbation responses with high correlation (e.g., correlation coefficient greater than 0.7), it was determined that they had a strong coupling with spectral anomalies.

[0108] The perturbation coupling strength value of each pixel is written into the corresponding image index to construct a complete dynamic perturbation factor map. This map describes the interference ability of each perturbation factor on spectral anomaly drift with pixel-level precision, providing input basis for subsequent weight adjustment and classification threshold adaptation of the recognition model.

[0109] One specific method for constructing a dynamic interference factor map is as follows:

[0110] The construction of the dynamic disturbance factor map is based on the spatial coupling strength between the spectral anomaly drift region and the disturbance factor, forming a two-dimensional distribution map with pixel-level spatial resolution. The construction process consists of the following specific steps:

[0111] In the registered image space, the perturbation parameters corresponding to each spectral anomaly drift region are analyzed pixel by pixel, including the concentration of suspended particulate matter (such as SPM), the concentration of dissolved organic matter (such as CDOM), and the incident light intensity perturbation value. For each image pixel, its perturbation parameter value and the perturbation change rate at the corresponding time point are extracted.

[0112] Then, based on the changes in the first derivative extracted from the spectral derivative curve (i.e., the trend of spectral shape change) and the changes in the aforementioned perturbation parameters, Pearson correlation analysis is performed. This analysis is conducted within a sliding window, linearly fitting each perturbation factor to the change in derivative curvature and recording its correlation coefficient. If the correlation coefficient exceeds a preset threshold (e.g., 0.6), it is determined that the perturbation factor has a significant coupling with the spectral drift of the current pixel.

[0113] For each perturbation factor, a coupling strength distribution map in the image space is generated. The coupling strength is defined as the correlation score between the perturbation response gradient value (i.e., the product of the perturbation value difference and the perturbation change rate) and the spectral curvature change value. Its value range can be normalized to between 0 and 1, and the larger the value, the higher the perturbation strength.

[0114] The coupling intensity distribution maps of all perturbation factors are weighted and superimposed to form a comprehensive dynamic perturbation factor map. The weight allocation can be set according to the empirical values ​​of the frequency and intensity of the influence of each perturbation factor on spectral anomaly drift in historical observation data. For example, if the CDOM perturbation dominates spectral anomaly drift in most scenarios, then the superposition weight of its corresponding coupling map is set to 0.5, SPM is set to 0.3, and light intensity is set to 0.2.

[0115] The constructed atlas is a two-dimensional matrix of the same size as the original image. Each pixel contains a disturbance factor value between 0 and 1, which is used to characterize the degree to which it is affected by background disturbance sources in the current environment. The atlas can be used for subsequent tasks such as model weight adjustment, adaptive classification threshold control, and disturbance robustness enhancement.

[0116] The calculation of perturbation coupling strength is performed on a per-pixel basis within the spectral anomaly drift region. First, for each pixel, the perturbation response gradient value corresponding to it in the perturbation mapping sequence is obtained. The perturbation response gradient value represents the product of the intensity change of the perturbation factor and the temporal gradient. This value is obtained by calculating the ratio of the perturbation factor's unit time difference to its intensity difference at that pixel's spatial location. For example, for wave perturbation, the response gradient value is the rate of change of wave height per unit time. The derivative spectrum trend corresponding to that pixel is extracted from the purified spectral feature set. The derivative spectrum trend represents the first derivative sequence of reflectance in a continuous band, with the unit being the change in reflectance per nanometer. The sliding window method is commonly used to extract continuous derivative changes. The window length can be set to 5 bands. The average derivative change within each window is calculated. When establishing correlation analysis, after aligning the perturbation response gradient sequence and the derivative spectrum change trend sequence in time, the Pearson correlation coefficient is calculated on each pixel. The correlation coefficient is defined as the ratio of the product of the covariance of the two sequences to their respective standard deviations. The value range is [-1, 1], which represents the strength and direction of the linear relationship between the two sequences.

[0117] The disturbance generation module establishes a pixel-level spatial coupling relationship between abnormal drift regions and disturbance factors, and constructs a dynamic disturbance factor map. This enables the system to dynamically sense and quantify the disturbance source intensity, effectively enhancing the system's ability to trace the cause of drift and analyze the response.

[0118] The adjustment and recognition module includes:

[0119] High-perturbation-intensity regions are extracted from the dynamic interference factor map. By setting a perturbation intensity threshold (e.g., regions higher than the 80th percentile in the map), significantly disturbed image pixels are identified. The spatial distribution pattern and frequency of these high-perturbation regions in the whole image are statistically analyzed. The distribution pattern can be quantified by cluster center density and dispersion to characterize the concentration and localization trend of the perturbation source.

[0120] Based on the statistically determined distribution of high-perturbation regions, the weight coefficients of the spectral recognition model in the image space are adjusted. Specifically, pixels in high-perturbation regions are assigned lower recognition model attention weights to reduce the interference of perturbation on the recognition results; while the weights of low-perturbation regions are maintained or increased to enhance the discriminative contribution of stable regions in the recognition process. For example, the weight coefficient for high-perturbation regions is set to 0.6, and that for low-perturbation regions is set to 1.2, achieving a differentiated response.

[0121] The purified spectral feature set is input into the current identification model to predict phytoplankton targets, and the predicted labels output by the model are obtained. Based on this, stable labels corresponding to the same region in historical samples are extracted, and the differences between the two are compared to calculate the confidence bias. The confidence bias is the absolute value of the difference between the predicted probability and the stable probability of the historical label, reflecting the reliability of the current identification result. For example, if the current model predicts that a pixel belongs to the "phytoplankton A" category with a probability of 0.72, while the historical stable sample probability is 0.91, then the confidence bias is 0.19.

[0122] Finally, the model classification threshold is dynamically adjusted based on the joint score of the confidence bias and the interference factor intensity. The joint score is calculated by multiplying the normalized confidence bias and interference intensity value of each pixel. The larger the joint score, the greater the interference affecting the pixel and the more unstable the model recognition. The system sets an adaptive adjustment range for the classification threshold based on the joint score. For example, if the joint score is greater than 0.15, the classification threshold for that pixel is increased from 0.5 to 0.6 to improve the conservatism of recognition of uncertain areas; if the score is less than 0.05, the threshold can be appropriately relaxed to improve recognition sensitivity.

[0123] The adjustment and identification module adaptively adjusts the identification model by fusing dynamic interference maps and the steady-state characteristics of purification spectra. It can automatically optimize the identification strategy according to interference conditions, achieve stable identification of plankton in multi-disturbance scenarios, and improve the system's environmental adaptability and classification accuracy.

[0124] The above formulas are all dimensionless calculations. The formulas are derived from software simulations using a large amount of collected data, and are the closest to the real situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0125] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0126] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0127] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0128] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0129] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A hyperspectral imaging detection system for marine plankton, characterized in that: It includes a marine water body data acquisition module, a spectral mapping module, a purification spectrum construction module, a spectral change analysis module, a disturbance generation module, and a regulation identification module, with each module connected by a signal. The marine water body data acquisition module is used to acquire hyperspectral image sequences and environmental disturbance parameters of the target marine water body, perform time-series synchronous calibration, and generate disturbance mapping sequences. The spectral mapping module is used to extract the spectral perturbation morphological features of the dominant bands of suspended particulate matter and dissolved organic matter in the perturbation mapping sequence, construct the perturbation contribution weight matrix and map it to the image pixel space; The purification spectrum construction module is used to perform pixel-level spectral stripping based on the perturbation contribution weight matrix, extract the purification spectral curve of the planktonic target region, and construct the purification spectral feature set. The spectral change analysis module is used to analyze the trend of derivative spectrum changes in the purified spectral feature set, and, in combination with local image texture intensity, locate regions of abnormal spectral drift. The disturbance generation module is used to establish the spatial coupling relationship between the spectral anomaly drift region and the background disturbance source of the water body, and generate a dynamic disturbance factor spectrum; The adjustment and identification module is used to adjust the weight distribution and target classification threshold of the spectral identification model according to the steady-state fit between the dynamic interference factor spectrum and the purification spectral feature set, so as to achieve adaptive planktonic identification under different interference conditions.

2. The hyperspectral imaging detection system for marine plankton according to claim 1, characterized in that: The marine water body data acquisition module includes: An integrated observation platform equipped with hyperspectral imaging and environmental disturbance parameter acquisition was deployed in the target sea area; Hyperspectral imaging acquires a sequence of surface water images at fixed time intervals and records the timestamp and geographic coordinates of each frame. The environmental disturbance parameter acquisition system collects real-time data on suspended particulate matter concentration, dissolved organic matter concentration, incident light intensity, and wave disturbance values, generating a disturbance parameter stream. The time synchronization controller aligns the image sequence with the disturbance parameter stream one by one using a unified time reference to construct a time synchronization structure; Each image pixel is combined with the perturbation parameters at the corresponding time point to form a perturbation mapping sequence.

3. The marine planktonic hyperspectral imaging detection system according to claim 2, characterized in that: The spectral mapping module includes: In the perturbation mapping sequence, spectral response analysis is performed on the perturbation parameters bound to each image pixel to extract the dominant bands of suspended particulate matter and dissolved organic matter; Based on the dominant band, the spectral shift value, band slope change value and reflectance fluctuation rate are calculated to construct the perturbation morphology feature vector; The perturbation morphological feature vector is normalized to generate the perturbation contribution ratio. The perturbation contribution ratio is mapped to the image cell index to construct a perturbation contribution weight matrix, which represents the distribution weight of each perturbation factor in the image space.

4. The hyperspectral imaging detection system for marine plankton according to claim 3, characterized in that: The purification spectrum building module includes: For each image pixel, a perturbation spectral estimation model is constructed using the perturbation contribution weight matrix, and the perturbation components are fitted to the original reflectance spectrum. The estimated perturbation spectrum is stripped from the original reflection spectrum band by band to output the purified spectrum; Intensity normalization and band alignment are performed on the purified spectrum to generate a standardized spectral curve; The standardized spectral curves of all pixels are used to form a purified spectral feature set.

5. The hyperspectral imaging detection system for marine plankton according to claim 4, characterized in that: The spectral variation analysis module includes: First-order derivative analysis and sliding curvature extraction are performed on each spectral curve in the purification spectral feature set to identify the peak shift position and the magnitude of change. By comparing the band positions in the curvature abrupt change range with the original peak positions, the drift bands and directions are extracted. Based on the curvature abrupt change and the texture intensity value of the corresponding extracted pixel, a consistency evaluation index between spectral derivative and texture response is constructed. The pixel region is analyzed based on the consistency evaluation index, and the spectral anomaly drift region of the pixel region is marked.

6. The hyperspectral imaging detection system for marine plankton according to claim 5, characterized in that: Based on the curvature abrupt change and the image texture intensity value of the corresponding pixel, a consistency evaluation index between the spectral derivative and the texture response is constructed, including: At the image pixel corresponding to the spectral drift band, the texture intensity value calculated by the gray-level co-occurrence matrix is ​​extracted and a mapping relationship is established with the curvature abrupt value. The consistency evaluation index is obtained by dividing the spectral curvature abrupt change value of the band corresponding to each image pixel by the image texture intensity value at the same location.

7. The marine plankton hyperspectral imaging detection system according to claim 6, characterized in that: The disturbance generation module includes: Establish a pixel-level registration relationship between the spatial coordinate map of the spectral anomaly drift region and the spatial distribution map of each perturbation factor in the perturbation mapping sequence; Calculate the difference in perturbation intensity and the relative rate of change between each drift pixel and the corresponding perturbation factor to form the perturbation response gradient value; The correlation between the perturbation response gradient value and the trend of the derivative spectrum is analyzed to calculate the perturbation coupling strength. The perturbation coupling strength is written into the image cell index to construct a dynamic perturbation factor map.

8. The hyperspectral imaging detection system for marine plankton according to claim 7, characterized in that: The perturbation coupling strength is obtained by calculating the Pearson correlation coefficient between the perturbation response gradient value of each image pixel in the spectral anomaly drift region and the trend of the derivative spectrum. The perturbation response gradient value is the change in the intensity of the perturbation factor per unit time, and the trend of the derivative spectrum is the first derivative sequence of the reflectance change in continuous bands.

9. A hyperspectral imaging detection system for marine plankton according to claim 8, characterized in that: The perturbation coupling strength is written under the image cell index to construct a dynamic perturbation factor map, including: Establish the registration relationship between the image pixel index of the spectral anomaly drift region and the spatial distribution map of each perturbation factor in the perturbation mapping sequence; The perturbation response gradient value is obtained by multiplying the perturbation intensity difference of each drift image pixel under the corresponding perturbation factor with the perturbation change rate. The perturbation coupling strength is obtained by calculating the correlation between the perturbation response gradient value and the change value of the first derivative in the spectral derivative curve within a sliding window. The perturbation coupling strength is written into the image cell index to generate a dynamic perturbation factor map, which is used to characterize the intensity of the perturbation source's influence on each cell.

10. A hyperspectral imaging detection system for marine plankton according to claim 9, characterized in that: The adjustment and recognition module includes: Distribution patterns and frequencies of high-disturbance-intensity regions in the statistical dynamic disturbance factor spectrum; Adjust the weight coefficients of the corresponding image regions in the planktonic identification model according to the interference intensity distribution; After performing model prediction on the purification spectral feature set, the confidence bias between the predicted label and the historical stable label is calculated; Based on the combined results of confidence bias and spatial interference factor, the classification threshold is dynamically updated to enhance target category recognition under perturbation conditions.

Images