Plant species accurate classification method using image data processing algorithm

Abstract

The application discloses a plant species accurate classification method using an image data processing algorithm, and relates to the technical field of plant species classification, and comprises the following steps: obtaining water body data and polarized multispectral image data of a target object; performing pixel polarization decomposition on the polarized multispectral image data to generate a polarization component image; performing pixel-by-pixel mapping on the water body data and the polarization component image to generate a scattering correction weight field; and performing coupling calculation on the scattering correction weight field and the polarization component image according to the pixel neighborhood gradient of the water body data to generate an initial compensation estimate. Through the introduction of the iteration mechanism controlled by the discrimination threshold, the scattering correction weight can be dynamically adjusted according to the polarization component change of the local area, the local compensation failure caused by the uneven spatial distribution of the water body turbidity is avoided, and through the iterative optimization of the fusion comparison result and the difference analysis data, the learner can be self-adapted to the water body conditions with different turbidity.

Description

Technical Field

[0001] This invention relates to the field of plant species classification technology, specifically to a method for accurate plant species classification using image data processing algorithms. Background Technology

[0002] With the development of remote sensing monitoring, environmental protection, and smart agriculture, the automatic identification and classification of plant species using image data processing algorithms has become an important direction in botanical research and aquatic ecological management. Especially in aquatic environments, the types and distribution of floating plants are directly related to ecosystem health, eutrophication levels, and water quality evolution trends. In recent years, multispectral and polarization imaging technologies have been gradually applied to plant identification. Their advantage lies in their ability to capture subtle plant features in both spectral and polarization dimensions, providing richer data sources for fine classification.

[0003] However, in the process of implementing the technical solution of the invention in the embodiments of this application, it was found that the above-mentioned technology has at least the following technical problems:

[0004] Existing plant classification methods include those based on spectral features. These methods typically rely on the differences between plants in the visible and near-infrared bands, classifying them by constructing vegetation indices or matching spectral curves. Their advantage lies in their intuitiveness and ease of implementation. However, in turbid water, the spectral features of plants are often masked by water absorption and reflection due to turbidity and scattering effects. Furthermore, the scattering and absorption effects of water exhibit high non-uniformity at the pixel scale, making it difficult to achieve fine-grained compensation for each pixel of the image, thus resulting in low accuracy of the generated plant classification results. Summary of the Invention

[0005] The purpose of this invention is to provide a method for accurate classification of plant species using image data processing algorithms, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, this invention discloses a method for accurate classification of plant species using image data processing algorithms, applied to the identification of species of floating plants in turbid water, comprising the following steps:

[0008] Acquire water body data and polarization multispectral image data of the target object:

[0009] Pixel polarization decomposition is performed on the polarization multispectral image data to generate polarization component images;

[0010] The water body data is mapped pixel-by-pixel to the polarization component image to generate a scattering correction weight field.

[0011] Based on the pixel neighborhood gradient of the water body data, the scattering correction weight field and the polarization component image are coupled and calculated to generate an initial compensation estimate.

[0012] The initial compensation estimate is applied to the polarization component image, and the residual of the polarization component image before and after compensation is calculated. The residual is used as a supervision signal to train the pre-constructed compensation correction model.

[0013] The polarization component images before and after compensation are input into the trained compensation and correction model, and the compensation and correction terms are output.

[0014] Using the scattering correction weight field as the fusion coefficient, the initial compensation estimate and the compensation correction term are weighted and fused pixel by pixel, and the weighted fusion result is applied to the polarization multispectral image data to generate a compensation image;

[0015] The compensated image is subjected to spectral correction and classification processing to generate plant classification results.

[0016] Secondly, this invention discloses a precise plant species classification system utilizing image data processing algorithms, comprising:

[0017] The data acquisition module is used to acquire water body data and polarization multispectral image data of the target object;

[0018] The data processing module is used to perform pixel polarization decomposition on the polarization multispectral image data to generate polarization component images;

[0019] The water body data is mapped pixel-by-pixel to the polarization component image to generate a scattering correction weight field.

[0020] The initial compensation generation module is used to perform coupled calculations on the scattering correction weight field and the polarization component image based on the pixel neighborhood gradient of the water body data to generate an initial compensation estimate.

[0021] The compensation correction term generation module is used to apply the initial compensation estimate to the polarization component image, calculate the difference between the polarization component images before and after compensation, and use the difference as a supervision signal to train the pre-constructed compensation correction model.

[0022] The compensated polarization component image is input into the trained compensation and correction model, and the compensation and correction term is output.

[0023] The compensation image generation module is used to use the scattering correction weight field as a fusion coefficient to perform weighted fusion of the initial compensation estimate and the compensation correction term pixel by pixel, and apply the weighted fusion result to the polarization multispectral image data to generate a compensation image.

[0024] The plant classification module is used to perform spectral correction and classification processing on the compensated image to generate plant classification results.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] 1. This scheme introduces an iterative mechanism with a discrimination threshold control, which can dynamically adjust the scattering correction weights according to the changes in polarization components in local areas. This avoids local compensation failure caused by uneven spatial distribution of water turbidity. By fusing comparison results and difference analysis data for iterative optimization, the learner can adapt to water conditions with different turbidity levels. Under the premise of ensuring computational efficiency, the scattering correction weights can accurately match the actual scattering characteristics of each pixel, providing a reliable data foundation for subsequent spectral correction and classification processing.

[0027] 2. This scheme combines polarization decomposition with scattering correction weight field, which preserves the physical constraints of water optical properties and captures local features using residual supervision mechanism, significantly improving the accuracy and reliability of the compensation process. The fusion mechanism of model and data-driven method ensures the physical rationality of the compensation results and enhances the adaptability of local scattering features, so that the generated compensation image can accurately reflect the true spectral characteristics of the plant. Attached Figure Description

[0028] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0029] Figure 1 This is a flowchart illustrating the steps of a method for accurate plant species classification using image data processing algorithms according to the present invention.

[0030] Figure 2 This is a schematic diagram of the process for generating the scattering correction weight field provided by the present invention;

[0031] Figure 3 This is a schematic diagram of the process for generating a stable reflectance spectrum estimate provided by the present invention;

[0032] Figure 4 This is a schematic diagram of the process for generating motion consistency scores provided by the present invention;

[0033] Figure 5 This is a schematic diagram illustrating the module functions of a precise plant species classification system utilizing image data processing algorithms provided by the present invention. Detailed Implementation

[0034] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0035] Application Overview:

[0036] In existing technologies, plant classification methods based on spectral features mainly rely on the spectral differences between the visible and near-infrared bands to achieve classification through vegetation indices or spectral curve matching. However, in turbid water environments, water turbidity and scattering effects can mask the spectral characteristics of floating plants, making it difficult for traditional methods to achieve pixel-by-pixel fine compensation. The uneven scattering and absorption effects of the water further exacerbate the distortion of spectral data, resulting in a significant decrease in classification accuracy.

[0037] To address the aforementioned issues, this study found that existing technologies do not adequately consider the coupling relationship between polarization information and water scattering, resulting in a lack of physical constraints in the compensation process. By analyzing the correlation between polarization multispectral data and water parameters, this study proposes mapping polarization decomposition components to water data pixel by pixel to establish a scattering correction weight field. Based on this, an initial compensation estimate is generated by combining neighborhood gradient information, and a correction model is trained through a residual supervision mechanism to form a dynamic compensation correction term. Finally, by fusing the physical model and data-driven results, the study achieves effective recovery of spectral features and improves the accuracy of plant classification results.

[0038] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0039] Example 1:

[0040] Please see Figure 1 A method for accurate plant species classification using image data processing algorithms, applied to the identification of floating plants in turbid water, includes the following steps:

[0041] Acquire water body data and polarization multispectral image data of the target object:

[0042] Pixel polarization decomposition is performed on polarization multispectral image data to generate polarization component images;

[0043] The water body data is mapped pixel by pixel to the polarization component image to generate a scattering correction weight field.

[0044] Based on the pixel neighborhood gradient of the water body data, the scattering correction weight field and polarization component image are coupled and calculated to generate an initial compensation estimate.

[0045] The initial compensation estimate is applied to the polarization component image, and the residuals of the polarization component images before and after compensation are calculated. The residuals are used as a supervision signal to train the pre-built compensation correction model.

[0046] Input the polarization component images before and after compensation into the trained compensation and correction model, and output the compensation and correction term;

[0047] Using the scattering correction weight field as the fusion coefficient, the initial compensation estimate and the compensation correction term are weighted and fused pixel by pixel, and the weighted fusion result is applied to polarization multispectral image data to generate a compensation image.

[0048] The compensated image is subjected to spectral correction and classification processing to generate plant classification results.

[0049] Among them, water body data refers to a set of structured information that reflects the optical, physical and chemical properties of water bodies in space and time;

[0050] Polarization multispectral image data refers to a collection of multi-channel image data that simultaneously contains spectral and polarization state information.

[0051] Pixel polarization decomposition refers to the separation and processing of polarization information in a multispectral image. For example, the Stokes vector decomposition method can be used to extract the intensity components of light under different polarization states.

[0052] Polarization component images refer to the image set obtained by pixel-level polarization decomposition of polarization multispectral images.

[0053] The scattering correction weight field refers to the spatial distribution matrix that reflects the scattering intensity of water bodies, for example, it is generated by calculating the correlation between water body data and polarization components;

[0054] Pixel neighborhood gradient refers to a type of numerical index used to describe the spatial variation characteristics of the local area surrounding each pixel in water body data.

[0055] Initial compensation estimation refers to an index used to correct for the effects of water scattering and absorption on multispectral image data;

[0056] Residual refers to the difference in the change of polarization component image before and after the application of initial compensation estimation;

[0057] Pre-built compensation correction models refer to machine learning models based on residual supervision, such as those using convolutional neural network structures;

[0058] The compensation adjustment term refers to the incremental adjustment to the initial compensation estimate;

[0059] Compensated images refer to the image results after scattering and absorption perturbation correction of polarization multispectral images in turbid water environments;

[0060] Spectral correction and classification processing refers to the process of further correcting the spectral information of each pixel in the compensated image, thereby identifying the plant category.

[0061] Plant classification results refer to pixel-level or target-level classification outputs used to identify specific species of floating plants in water.

[0062] This scheme, through the combined application of polarization decomposition and scattering correction weight field, not only preserves the physical constraints of the optical properties of water bodies but also captures local features using a residual supervision mechanism, effectively distinguishing between plant body reflection and water surface reflection, thus overcoming the shortcomings of traditional spectroscopic methods that are susceptible to water interference. Through the above technical solution, this application significantly improves the accuracy and reliability of the compensation process, providing a high-quality data foundation for subsequent classification.

[0063] The above describes a complete scheme for accurate plant species classification using image data processing algorithms. The following section describes how to acquire water body data and polarization multispectral image data of the target object, specifically including:

[0064] Acquire initial water body data and initial polarization multispectral image data of the target object;

[0065] The initial polarization multispectral image data is compared frame by frame with the initial water body data corresponding to the time and space, and the initial polarization multispectral image data is corrected according to the frame-by-frame comparison results to generate polarization multispectral image data.

[0066] The polarization multispectral image data is inverted, and the difference between the polarization multispectral image data before and after the inversion is analyzed. The difference analysis results are used as constraints to correct the initial water body data and generate water body data.

[0067] Among them, initial water body data refers to raw physical measurement data directly collected from the aquatic environment;

[0068] Initial polarization multispectral image data refers to the raw multispectral image information obtained from the target water body area;

[0069] Frame-by-frame comparison results refer to the set of quantitative and qualitative information obtained by comparing the initial polarization multispectral image data with the initial water body data at the corresponding time frame and location at the pixel level and region level.

[0070] Inversion refers to the process of calculating the optical parameters of water bodies based on polarization multispectral image data;

[0071] The difference analysis results refer to the generation of an error distribution map by comparing polarization multispectral image data before and after inversion. Specifically, this can be achieved using pixel-level difference operations or statistical analysis methods.

[0072] The above content will be described in detail below:

[0073] The system acquires initial water body data for the target area. This initial water body data is collected in real time by an array of sensors deployed on and underwater around the water surface. This data includes, but is not limited to, parameters such as optical transmittance, scattering coefficient, refractive index, and local turbidity distribution of the water body.

[0074] Acquire initial polarization multispectral image data of the target object. The initial polarization multispectral image data is acquired simultaneously in multiple bands and at different polarization angles by a multispectral imaging device equipped with a polarization filter. The data includes, but is not limited to, parameters such as image brightness distribution, polarization angle change and spectral intensity sequence.

[0075] The initial water body data and initial polarization multispectral image data are aligned and fused under a unified time axis and spatial reference coordinates;

[0076] The initial polarization multispectral image data is compared frame by frame with the corresponding spatiotemporal initial water body data:

[0077] Water body data intervals corresponding to the initial polarization multispectral image data in time and space are extracted from the water body data, and frame-by-frame comparison is performed with the initial polarization multispectral image data. During the frame-by-frame comparison stage, the image brightness distribution values ​​within the initial polarization multispectral image data are analyzed. polarization angle value With spectral intensity And so on, comparing each with the turbidity within the corresponding water body data range. absorption coefficient and scattering characteristic parameters By performing joint comparisons and establishing a two-way mapping function between timestamps and spatial coordinates, the offset residuals between polarization multispectral image data and water body data in the time and spatial domains are calculated. This is used as the frame-by-frame comparison result, and the specific calculation formula is as follows:

[0078]

[0079] In the formula, Indicates time and spatial location Offset residuals below Indicates time and spatial location Brightness distribution values ​​under, Indicates time and spatial location The turbidity below, Indicates time and spatial location The polarization angle value below, Indicates time and spatial location The absorption coefficient below, Indicates time and spatial location Spectral intensity below, Indicates time and spatial location The scattering characteristic parameters under the following conditions , and The above data have been normalized before calculation.

[0080] The initial polarization multispectral image data is superimposed pixel by pixel based on the frame-by-frame comparison results to generate polarization multispectral image data.

[0081] The polarization multispectral image data was inverted, and a difference analysis was performed on the polarization multispectral image data before and after the inversion. The results of the difference analysis were used as constraints to correct the initial water body data, generating water body data.

[0082] Water body spectral response predictions are generated from polarization multispectral image data using an inversion algorithm. A difference analysis is then performed between the predicted predictions and the polarization multispectral image data to form image inversion residuals. These residuals are used as the result of the difference analysis and as constraint terms to iteratively correct the initial water body data.

[0083] In each iteration, the initial water data of the corresponding pixels is adjusted according to the magnitude and direction of the image inversion residual until the image inversion residual converges and water data is generated.

[0084] This scheme, through a two-way correction mechanism of frame-by-frame comparison and inversion difference analysis, not only corrects the spectral distortion of polarization multispectral image data but also optimizes the physical accuracy of water body data, thus providing a more reliable data foundation for subsequent scattering compensation. Through the above technical solution, this application can effectively eliminate the problem of mismatch between polarization images and water body parameters caused by insufficient accuracy of data acquisition equipment or environmental interference, improve the spatiotemporal consistency of initial data, and thus improve the stability of plant spectral feature extraction, providing data assurance for improving classification accuracy.

[0085] The above describes the acquisition of water body data and polarization multispectral image data of the target object. The following describes pixel polarization decomposition of the polarization multispectral image data to generate polarization component images, specifically including:

[0086] Using polarization multispectral images as input, each pixel location has observations at different polarization angles. These observations are themselves the response sequence of the pixel in the polarization angle domain. At the same time, the polarization vector sequence of the pixel in the local spatial neighborhood (e.g., a 3×3 or 5×5 pixel window centered on the pixel) and the temporal neighborhood (e.g., consecutive frames of images) is obtained and used as context input.

[0087] In the processing, the response sequence of the pixel in the polarization angle domain is first normalized and noise filtered to generate a preliminary pixel vector. Then, the preliminary pixel vector is fused with the polarization vector sequence of the spatial neighborhood and the polarization vector sequence of the temporal neighborhood to generate the polarization component image of each pixel.

[0088] This scheme enhances the polarization information representation of pixels by normalizing and filtering the response sequence of each pixel in the polarization angular domain, and fusing the polarization vector sequences of its spatial and temporal neighborhoods, while maintaining spatial continuity and temporal consistency. This generates more accurate, stable, and interpretable polarization component images, providing reliable input data for subsequent water spectral inversion, scattering correction, and plant classification, thereby improving overall recognition accuracy and anti-interference capability.

[0089] The above describes pixel-by-pixel polarization decomposition of polarization multispectral image data to generate polarization component images. The following describes pixel-by-pixel mapping of water body data to polarization component images to generate a scattering correction weight field. Please refer to [link / reference]. Figure 2 , Figure 2 This is a schematic diagram of the process for generating a scattering correction weight field provided in an embodiment of this application. Generating the scattering correction weight field specifically includes:

[0090] Water body data and differential analysis results are input into a pre-built learner to generate initial scattering correction weights for each pixel, which are then applied to the polarization component image and compared with the polarization component image.

[0091] If the comparison result is greater than the preset discrimination threshold, the initial scattering correction weight is used as the scattering correction weight of the corresponding pixel, and then a scattering correction weight field is generated.

[0092] Otherwise, the comparison results and difference analysis results are merged and input into a pre-built learner along with the water body data to iteratively generate initial scattering correction weights until the preset discrimination threshold is met, thereby generating a scattering correction weight field.

[0093] The learner refers to a computational model that can generate pixel-level scattering correction weights based on input data, which can be implemented using a multilayer perceptron or a convolutional neural network.

[0094] Initial scattering correction weights refer to the numerical coefficients used to perform preliminary scattering compensation on polarization component image pixels.

[0095] The comparison result refers to a quantitative indicator used to measure the difference between the correction effect produced by the initial scattering correction weight of each pixel when correcting the polarization component image and the original polarization component image.

[0096] The discrimination threshold is a critical value used to determine whether the initial scattering correction weights meet the accuracy requirements. Specifically, it can be obtained by statistically analyzing the pixel residuals using historical observation data and calibration samples, and calculating the mean and standard deviation of the residual distribution under different lighting conditions, water turbidity, and scattering characteristics. Then, the mean of the residuals is added to the standard deviation by a certain multiple to obtain the value.

[0097] Iterative generation refers to the process of recalculating the initial scattering correction weights by adjusting the input parameters multiple times. Specifically, it can be implemented using gradient descent or genetic algorithms.

[0098] The above content will be described in detail below:

[0099] Construct a learner; the learner includes an input layer, a feature extraction layer, a non-linear mapping layer, and an output layer;

[0100] The input layer receives water body data and differential analysis results, and performs standardization and normalization on them to form preliminary pixel feature vectors.

[0101] The feature extraction layer performs high-dimensional feature extraction on the initial pixel feature vector;

[0102] The nonlinear mapping layer maps the high-dimensional feature extraction results through a nonlinear activation function (such as ReLU or Sigmoid) to generate the initial scattering correction weights for each pixel.

[0103] The output layer outputs the initial scattering correction weights for each pixel;

[0104] The water body data and difference analysis results are input into a pre-built learner to generate an initial scattering correction weight for each pixel. This weight is then multiplied by the polarization component image and the difference between the two images is calculated to obtain the comparison results.

[0105] If the comparison result is greater than the preset discrimination threshold, the initial scattering correction weight is used as the scattering correction weight of the corresponding pixel. The scattering correction weights of all pixels are combined to generate a scattering correction weight field.

[0106] Otherwise, the comparison results and difference analysis results are fused and input into a pre-built learner along with the water body data. Initial scattering correction weights are iteratively generated until a preset discrimination threshold is met. Finally, the scattering correction weights of all pixels are combined to generate a scattering correction weight field. Its expression is as follows:

[0107]

[0108] In the formula, Represents the scattering correction weight field. Indicates the first Line number Scattering correction weights for column pixels.

[0109] This scheme introduces an iterative mechanism with a discrimination threshold control, which dynamically adjusts the scattering correction weights based on changes in the polarization components of local regions. By fusing comparison results and difference analysis data for iterative optimization, the learner can adapt to water conditions with different turbidity levels. Through the above technical solution, this application solves the problem of insufficient accuracy at the pixel scale in traditional scattering correction methods, avoids local compensation failure caused by uneven spatial distribution of water turbidity, and, through dynamic discrimination and iterative optimization mechanisms, ensures that the scattering correction weights accurately match the actual scattering characteristics of each pixel while maintaining computational efficiency, providing a reliable data foundation for subsequent spectral correction and classification processing.

[0110] The above describes the pixel-by-pixel mapping of water body data to polarization component imagery to generate a scattering correction weight field. The following describes the coupling calculation of the scattering correction weight field and polarization component imagery based on the pixel neighborhood gradient of the water body data to generate an initial compensation estimate, specifically including:

[0111] Based on the pixel neighborhood gradient of the water body data, the scattering correction weight field and polarization component image are coupled and calculated to generate an initial compensation estimate. The specific calculation formula is as follows:

[0112]

[0113] In the formula, Represents pixels The initial compensation estimate, Represents pixels in polarization component image The vector value, Represents the pixel in the scattering correction weight field Scattering correction weights, Represents pixels Pixel neighborhood gradient of water body data This represents the coupling function. All the data above have been normalized before calculation.

[0114] Based on the initial compensation estimate, the polarization component image is weighted and corrected, and the residual of the polarization component image before and after compensation is obtained by difference calculation.

[0115] This scheme, through the coupled calculation of pixel neighborhood gradient and scattering correction weight and the residual generation method, can achieve fine compensation of polarization component images at the pixel scale, effectively correcting local spectral deviations caused by water scattering and absorption, while preserving spatial continuity and temporal consistency, thus providing a more reliable and stable spectral basis for the accurate classification of floating plants in turbid water.

[0116] The above describes the coupling calculation of the scattering correction weight field and polarization component image based on the pixel neighborhood gradient of the water body data to generate an initial compensation estimate. The following describes the training of the pre-built compensation correction model using the residual as a supervision signal, specifically including:

[0117] The scattering correction weight field and the residual are weighted and calculated to generate a weighted residual.

[0118] The similarity between the residual and the residual of the pixel in its spatial neighborhood is calculated and coupled with the weighted residual. The coupling result is used as a supervision signal to train the pre-built compensation correction model.

[0119] Among them, the weighted residual refers to the result of multiplying the scattering correction weight field with the residual pixel by pixel, which can be achieved by matrix dot multiplication.

[0120] Similarity refers to the correlation measure between a residual and the residuals of its neighboring pixels, which can be achieved by using cosine similarity or Euclidean distance.

[0121] The coupling result refers to the signal obtained by linearly combining or nonlinearly fusing similarity and weighted residuals, which can be achieved by using weighted summation or neural network fusion modules.

[0122] The above content will be described in detail below:

[0123] The scattering correction weight field and the residual are weighted and calculated to generate a weighted residual.

[0124] Calculate residuals The residual of its spatial neighborhood pixels The similarity is calculated using the following formula:

[0125]

[0126] In the formula, Represents pixels similarity, Represents pixels The spatial neighborhood set, Represents pixels in spatial neighborhood The residual, Represents pixels The residual, This represents the residual similarity function, such as cosine similarity or normalized Euclidean distance. All the above data have been normalized before calculation.

[0127] The model is then coupled with the weighted residuals through a weighted calculation, and the coupling result is used as a supervision signal to train the pre-built compensation correction model.

[0128] Construct a compensation and correction model; the compensation and correction model includes constructing an input layer, a convolutional feature extraction layer, a spatiotemporal coupling layer, a nonlinear mapping compression layer, and an output layer.

[0129] The input layer of the compensation correction model is constructed to receive the polarization component images before and after compensation, and stacked according to the corresponding positions of pixels to form a dual-channel feature vector for each pixel. The dual-channel feature vector of each pixel is normalized and noise filtered to form a preliminary feature vector matrix corresponding to each pixel.

[0130] The convolutional feature extraction layer uses small-sized convolutional kernels (such as 3×3 or 5×5) to extract local spatial features of the initial feature vector matrix. At the same time, through channel blending, the dual-channel feature vector of each pixel is mapped to a high-dimensional feature space. Combined with local spatial features, a high-dimensional feature tensor containing local spatial features and channel information is generated.

[0131] The spatiotemporal coupling layer calculates the similarity between the high-dimensional feature tensor of each pixel and the features of its spatial and temporal neighbors. Specifically, it can use weighted similarity calculation and feature fusion to integrate neighborhood features into the high-dimensional feature tensor of the current pixel through weighted summation or attention mechanism, thereby generating a coupled feature tensor containing spatiotemporal coupling information.

[0132] The nonlinear mapping compression layer maps the coupled feature tensor to features through a pixel-wise nonlinear activation function (such as ReLU or Leaky ReLU), and compresses the channel dimension through a 1×1 convolution or fully connected layer to form a modified feature vector for each pixel.

[0133] The compensation correction model output layer generates and outputs compensation correction terms by performing a pixel-by-pixel linear mapping on the correction feature vector;

[0134] During the loss calculation stage, each compensation correction term output by the compensation correction model is compared pixel by pixel with the coupling result to form a residual tensor.

[0135] To fully utilize the spatial and temporal correlations between pixels, the loss function design incorporates multiple combinations based on the residual tensor, introducing three types of regularization constraints:

[0136] Spatial Consistency Regularization It is obtained by calculating the residual difference between each pixel and its neighboring pixels;

[0137] Temporal smoothing regularization This was obtained by analyzing the changes in pixel residuals across consecutive frames;

[0138] Water data consistency regularization The compensation correction item is obtained by comparing it with the water body data;

[0139] The specific expression for the loss function is as follows:

[0140]

[0141] In the formula, Indicates the total loss. Represents pixels Compensation correction items, Represents pixels The coupling result, Indicates the weighting factor. Indicates space consistency regularity, Indicates time-series smoothing regularization. Indicates a regularization for water body data consistency. , and This represents the corresponding weighting coefficient. All the above data have been normalized before calculation.

[0142] During the backpropagation phase, the gradient derivative of the weights of each layer in the compensation and correction model is calculated according to the designed loss function to form a gradient tensor. The update of each weight is corrected by combining the learning rate with the optimization algorithm (such as Adam or SGD) so that the output of the compensation and correction model gradually approaches the coupling result, while satisfying the spatial, temporal and physical consistency constraints.

[0143] During the iterative optimization phase, the forward propagation, loss calculation, and backpropagation processes are repeated multiple times. In each iteration, the compensation correction model continuously adjusts its parameters to make the output compensation correction terms smoother in space and time and more consistent with physical water body data. At the same time, it is sensitive to disturbances in key scattering regions, thereby completing the training of the compensation correction model.

[0144] This scheme effectively suppresses local overfitting caused by uneven scattering distribution by introducing a weighted residual and similarity coupling mechanism. At the same time, it enhances the model's constraint on neighborhood consistency and avoids interference from isolated noise points in the training process. Through the above technical solution, this application solves the problem of excessive noise in the supervision signal and unstable model training caused by water scattering effect in the prior art. By fusing spatial neighborhood similarity and residual information after weight adjustment, the robustness of the supervision signal is improved, enabling the compensation and correction model to more accurately capture the spatial variation law of scattering interference, thereby improving the accuracy of subsequent image compensation and classification.

[0145] The above describes training a pre-built compensation correction model using residuals as supervision signals. The following describes the generation of compensation images, specifically including:

[0146] Input the polarization component images before and after compensation into the trained compensation and correction model, and output the compensation and correction term;

[0147] Using the scattering correction weight field as the fusion coefficient, the initial compensation estimate and the compensation correction term are weighted and fused pixel by pixel, and the weighted fusion result is superimposed on the polarization multispectral image data pixel by pixel to generate the compensation image.

[0148] The above describes the generation of the compensated image. The following section describes the spectral correction and classification processing of the compensated image to generate plant classification results, specifically including:

[0149] Spectral fitting is performed on the compensated image and polarization component image to generate a stable reflectance spectrum estimate;

[0150] Cluster analysis is performed on the stable reflection spectrum estimates to generate a time series set of target spectra;

[0151] The target spectral time series set and the compensated image are input into a pre-constructed fusion discriminator, which outputs plant classification results.

[0152] Among them, spectral fitting refers to the process of matching the spectral data of the compensated image with the polarization component image through mathematical methods to eliminate spectral line shifts caused by scattering or noise.

[0153] Stable reflectance spectrum estimation refers to physically consistent spectral reflectance data obtained after correction processing. Specifically, it can be obtained by performing physical consistency correction on the compensated image and combining it with spectral residual correction.

[0154] Cluster analysis refers to classifying pixels into different categories based on the similarity of spectral features. Specifically, it can be implemented using the K-means algorithm or hierarchical clustering algorithm.

[0155] The target spectrum time series set refers to the set of spectral data of pixels within the same candidate spectrum cluster at different time points, which can be generated by weighting the candidate spectrum clusters.

[0156] A fusion discriminator is a machine learning model that integrates data from multiple sources to make classification decisions. Specifically, it can be implemented using a multi-branch neural network structure, with each branch corresponding to a different data processing strategy.

[0157] The above content will be described in detail below:

[0158] Physical consistency correction is performed on the compensated image based on the scattering correction weight field and water body data to obtain the reference spectral value of each pixel. The specific calculation formula is as follows:

[0159]

[0160] In the formula, Represents pixels Reference spectral values, Represents pixels Compensation image pixel values, Represents pixels Scattering correction weights, Indicates water body data in pixels The spectral values ​​at the location are all normalized before calculation.

[0161] The compensated image and polarization component image are spectrally fitted to generate initial spectral estimates for each pixel. The specific calculation formula is as follows:

[0162]

[0163] In the formula, Represents pixels Initial spectral estimation, Represents pixels In time Polarization component image at time, The spectral fitting function is represented by the function, and the reference spectral value is subtracted from it to generate the spectral residual. All the above data are normalized before calculation.

[0164] The temporal attention weight field is multiplied with the spectral residual, and the multiplication result is superimposed with the initial spectral estimate to generate a stable reflectance spectrum estimate for each pixel.

[0165] The temporal attention weight field is obtained by calculating the sum of squares of the spectral residuals of each pixel at different time points;

[0166] Cluster analysis was performed on the stable reflectance spectrum estimates to form candidate spectral clusters:

[0167] For each pixel's stable reflectance spectrum estimate, its similarity to the stable reflectance spectrum estimates of other pixels is calculated using the following formula:

[0168]

[0169] In the formula, Represents pixels Stable reflection spectrum estimation and pixel The similarity between the stable reflectance spectrum estimates, Represents pixels Stable reflection spectrum estimation, Represents pixels Stable reflection spectrum estimation, This represents the maximum value in the stable reflection spectrum estimation. This represents the minimum value in the stable reflectance spectrum estimation. All the above data have been normalized before calculation.

[0170] Clustering algorithms (such as hierarchical clustering, k-means clustering, or density clustering) are used to group similarity and generate candidate spectral clusters;

[0171] The temporal attention weight field is weighted with the stable reflectance spectrum estimate of each pixel in the candidate spectral cluster to generate the target spectral temporal set for each candidate spectral cluster.

[0172] Based on the pixels contained in each candidate spectral cluster, target detection and segmentation are performed on the compensated image to generate the pixel motion trajectory for each candidate spectral cluster:

[0173] In the target detection stage, based on the pixels contained in each candidate spectral cluster, the boundary of each candidate spectral cluster is identified from the compensated image using a pixel-level classification method with threshold segmentation.

[0174] In the segmentation process, pixels within candidate spectral clusters are distinguished from the background of the compensation image to form independent pixel sets. The position of each pixel in consecutive frames of the compensation image is tracked to obtain the pixel motion trajectory of each pixel within each candidate spectral cluster.

[0175] Based on the pixels contained in each candidate spectral cluster, extract water body sub-data with corresponding spatial location and time from the water body data;

[0176] pixel motion trajectory With corresponding water body sub-data The comparisons are performed to generate a motion consistency score for each candidate spectral cluster. The specific calculation formula is as follows:

[0177]

[0178] In the formula, Represents candidate spectral clusters Consistency score, Represents candidate spectral clusters The total number of pixels within, Represents pixels In time The pixel motion trajectory, Represents pixels In time Water body sub-data, Represents pixels The weight, This represents the motion consistency matching function, used to calculate the similarity between pixel trajectories and water flow fields. Cosine similarity or vector inner product can be used. All the above data have been normalized before calculation.

[0179] Construct a fusion discriminator; the fusion discriminator includes a fusion discriminator input layer, a data processing branch layer, a weighted fusion layer, and a fusion discriminator output layer;

[0180] The input layer of the fusion discriminator is used to receive the target spectrum temporal set, motion consistency score, temporal attention weight field and compensation image, and perform normalization processing on each of them;

[0181] The data processing branch layer consists of three main branches, each focusing on a different discrimination strategy and outputting its own fitness score:

[0182] Spectrum-dominated branch: Taking the target spectral time series set as input, it calculates the vector similarity of each candidate spectral cluster in the high-dimensional spectral space, and adjusts the temporal domain reliability by combining temporal attention weights, outputting a spectral feature fit vector. The specific calculation formula is as follows:

[0183]

[0184] In the formula, Represents candidate spectral clusters Spectral feature fit vector, Indicates the total number of time frames. Indicates time Temporal attention weights, Represents candidate spectral clusters Inner pixel In time spectral vector at time, Represents candidate spectral clusters In time The central spectral vector at time, The vector cosine similarity function is represented by the above data, which have all been normalized before calculation.

[0185] Motion-dominant branch: The motion consistency score is weighted according to the temporal attention weight to generate a motion reliability weight vector;

[0186] Polarization-dominated branch: Taking the temporal attention weight field and the compensation image as input, it performs residual correction and reliability assessment, and outputs a polarization consistency weight vector. The specific calculation formula is as follows:

[0187]

[0188] In the formula, Represents the polarization consistency weight vector. Represents the temporal attention weight field. Represents the compensated image. This represents the compensated polarization component image. This indicates element-wise multiplication. This represents the normalization function. All the data above have been normalized before calculation.

[0189] The weighted fusion layer is used to weight and fuse the spectral feature fit vector, motion reliability weight vector, and polarization consistency weight vector to generate the fusion feature vector of each candidate spectral cluster.

[0190] The output layer of the fusion discriminator matches the fusion feature vectors of each candidate spectral cluster with a predefined set of plant classification labels to generate the probability distribution of each candidate spectral cluster in each plant category. The specific calculation formula is as follows:

[0191]

[0192] In the formula, Represents candidate spectral clusters Belongs to the plant category The probability, Indicates the total number of plant categories. Candidate spectral clusters The fused feature vector, Indicate plant category Weighting factors Indicate plant category The bias term, all the above data have been normalized before calculation;

[0193] The plant category corresponding to the maximum probability of each candidate cluster is selected as the plant classification result of the corresponding candidate cluster and output.

[0194] Among them, the predefined plant classification label set is generated by normalizing and encoding the spectral characteristics of plants collected from system surveys and samples, generating unique corresponding category labels, and combining spectral characteristics, category labels and plant species;

[0195] The target spectrum temporal set, motion consistency score, temporal attention weight field, and compensated intermediate image are input into a pre-constructed fusion discriminator to output plant classification results.

[0196] This scheme generates stable reflectance spectrum estimates through spectral fitting and residual correction, effectively suppressing scattering noise; enhances the spectral distinguishability of different plant species through cluster analysis and time series set construction; and improves classification robustness in complex environments by integrating multidimensional information through a fusion discriminator. Through the above technical solutions, this application can reduce the interference of water turbidity on spectral features, accurately extract the stable reflectance characteristics of plants in polarization multispectral data, and significantly improve the accuracy of identifying floating plant species in turbid water by combining time series variation features and spatial distribution information.

[0197] The above describes the spectral correction and classification processing of the compensated image to generate plant classification results. The following describes the spectral fitting of the compensated image and polarization component image to generate a stable reflectance spectrum estimate. Please refer to [link / reference]. Figure 3 , Figure 3 This is a schematic diagram of the process for generating a stable reflectance spectrum estimate provided in an embodiment of this application. Generating a stable reflectance spectrum estimate specifically includes:

[0198] The physical consistency of the compensated image is corrected based on the scattering correction weight field and water body data to obtain the reference spectral value of each pixel;

[0199] The compensated image and polarization component image are spectrally fitted to generate an initial spectral estimate for each pixel, and the reference spectral value is subtracted from it to generate the spectral residual.

[0200] The temporal attention weight field is multiplied with the spectral residual, and the multiplication result is superimposed with the initial spectral estimate to generate a stable reflectance spectrum estimate for each pixel.

[0201] The temporal attention weight field is obtained by calculating the sum of squares of the spectral residuals of each pixel at different time points.

[0202] Among them, physical consistency correction refers to adjusting the compensation image based on a physical model through water body parameters and scattering correction weights, which can be achieved by combining the radiative transfer equation with the scattering coefficient matrix.

[0203] The spectral residual refers to the difference between the initial spectral estimate and the reference spectral value, which can be calculated using the least squares method.

[0204] Temporal attention weight field refers to a dynamic weight matrix generated based on the sum of squared spectral residuals at different time points.

[0205] This part has already been described in detail above, so I will not repeat it here.

[0206] This scheme introduces a temporal attention weight field, which can dynamically adjust the weights according to the temporal statistical characteristics of the spectral residuals. Combined with the superposition correction mechanism, it can effectively suppress noise interference while preserving the initial spectral features, thereby improving spectral stability. Through the above technical solution, this application can reduce the impact of dynamic scattering of water bodies on spectral estimation. By dynamically adjusting the residual weights through the temporal attention mechanism, combined with physical consistency correction and superposition correction, it can significantly improve the accuracy of stable reflectance spectrum estimation, providing a more reliable spectral feature basis for subsequent plant classification.

[0207] The above describes spectral fitting of the compensated image and polarization component image to generate a stable reflectance spectrum estimate. The following describes cluster analysis of the stable reflectance spectrum estimate to generate a target spectrum time series set, specifically including:

[0208] Cluster analysis is performed on the stable reflectance spectrum estimates to form candidate spectral clusters;

[0209] The temporal attention weight field is weighted with the stable reflectance spectrum estimate of each pixel in the candidate spectral cluster to generate the target spectral temporal set for each candidate spectral cluster.

[0210] Among them, the candidate spectral cluster refers to the set of pixels with similar spectral features formed by grouping the stable reflectance spectrum estimates through a clustering algorithm. Specifically, it can be implemented using K-means or hierarchical clustering algorithms.

[0211] This part has already been described in detail above, so I will not repeat it here.

[0212] This scheme introduces a temporal attention weight field, which combines the weight allocation of the time dimension with clustering, enabling the target spectral temporal set to dynamically capture the evolution of spectral features over time. Through the above technical solution, this application can effectively reduce the impact of temporal spectral data fluctuations caused by water turbidity on classification. By strengthening spectral features through weighted processing, the generated candidate spectral clusters have higher spectral consistency, thereby improving the accuracy of subsequent plant species differentiation.

[0213] The above describes cluster analysis for stable reflection spectrum estimation to generate target spectrum time series sets. The following section describes how, after generating the target spectrum time series sets for each candidate spectrum cluster, a motion consistency score is calculated for each candidate spectrum cluster. Please refer to [link / reference]. Figure 4 , Figure 4 This is a schematic diagram of the process for generating a motion consistency score according to an embodiment of this application. Generating a motion consistency score specifically includes:

[0214] Based on the pixels contained in each candidate spectral cluster, target detection and segmentation are performed on the compensated image to generate the pixel motion trajectory of each candidate spectral cluster.

[0215] Based on the pixels contained in each candidate spectral cluster, extract the corresponding water body sub-data from the water body data;

[0216] The pixel motion trajectory is compared with the corresponding water body sub-data to generate a motion consistency score for each candidate spectral cluster.

[0217] Among them, target detection and segmentation refers to the process of locating, identifying and dividing the pixels contained in the candidate spectral clusters in the spatial and spectral domains of the compensated image.

[0218] Pixel motion trajectory refers to the spatial position change path of pixels within a candidate spectral cluster at consecutive time points, which can be generated by optical flow or target tracking algorithm.

[0219] Water body subdata refers to the set of water body parameters associated with the spatial location of candidate spectral clusters, which may include parameters such as turbidity, flow velocity, and temperature.

[0220] Motion consistency score is a quantitative indicator of the degree of matching between pixel motion trajectory and flow velocity direction in water body sub-data. Specifically, it can be calculated using the cosine similarity between trajectory direction and flow velocity direction.

[0221] This part has already been described in detail above, so I will not repeat it here.

[0222] This solution introduces a motion consistency scoring mechanism to effectively distinguish between real plant targets and interference objects, solving the misclassification problem caused by neglecting dynamic characteristics in traditional methods. Through the above technical solution, this application can combine plant spectral characteristics and motion characteristics to eliminate the interference of the dynamic water environment on the classification results. The motion consistency scoring mechanism can effectively identify abnormal moving targets caused by water turbulence or foreign object interference, improve the ability to distinguish between floating plants and non-plant targets, and thus achieve more accurate plant species identification in turbid water scenarios.

[0223] As described above, after generating the target spectral time series set for each candidate spectral cluster, the process also includes calculating the motion consistency score for each candidate spectral cluster. The following describes how the target spectral time series set and the compensated image are input into a pre-constructed fusion discriminator to output plant classification results, specifically including:

[0224] The target spectrum temporal set, motion consistency score, temporal attention weight field and compensated intermediate image are input into a pre-constructed fusion discriminator to output plant classification results;

[0225] The pre-built fusion discriminator contains different data processing branches, each corresponding to a different discrimination strategy.

[0226] This part has already been described in detail above, so I will not repeat it here.

[0227] This scheme constructs a multi-branch fusion discriminator to simultaneously analyze features in three dimensions: spectral temporal variation, motion pattern consistency, and temporal stability. It enhances classification robustness by utilizing the complementarity between different discrimination strategies. Through the above technical solution, this application can effectively solve the problem of plant spectral feature distortion caused by scattering effects in turbid water, significantly improving classification accuracy. By fusing multi-dimensional information such as spectral temporal features, motion pattern consistency, and temporal stability, it can achieve accurate identification of floating plant species in complex water environments. In particular, when spectral features are severely interfered with, misjudgment can be avoided through joint analysis of motion patterns and spatiotemporal stability features.

[0228] Example 2:

[0229] Please see Figure 5 A precise plant species classification system utilizing image data processing algorithms, comprising:

[0230] The data acquisition module is used to acquire water body data and polarization multispectral image data of the target object;

[0231] The data processing module is used to perform pixel polarization decomposition on polarization multispectral image data to generate polarization component images;

[0232] The water body data is mapped pixel by pixel to the polarization component image to generate a scattering correction weight field.

[0233] The initial compensation generation module is used to perform coupled calculations on the scattering correction weight field and polarization component image based on the pixel neighborhood gradient of the water body data to generate an initial compensation estimate.

[0234] The compensation correction term generation module is used to apply the initial compensation estimate to the polarization component image and calculate the difference between the polarization component images before and after compensation. The difference is used as a supervision signal to train the pre-built compensation correction model.

[0235] The compensated polarization component image is input into the trained compensation and correction model, and the compensation and correction term is output.

[0236] The compensated image generation module is used to perform weighted fusion of the initial compensation estimate and the compensation correction term pixel by pixel using the scattering correction weight field as the fusion coefficient, and apply the weighted fusion result to the polarization multispectral image data to generate a compensated image.

[0237] The plant classification module is used to perform spectral correction and classification processing on the compensated images to generate plant classification results.

[0238] This embodiment has the same technical effects as Embodiment 1.

[0239] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. The data mentioned in this application, when used for calculations, have undergone normalization and other preprocessing to achieve dimensional uniformity.

[0240] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for accurate classification of plant species using image data processing algorithms, applied to the identification of floating plants in turbid water, characterized in that, Includes the following steps: Acquire water body data and polarization multispectral image data of the target object: Pixel polarization decomposition is performed on the polarization multispectral image data to generate polarization component images; The water body data is mapped pixel-by-pixel to the polarization component image to generate a scattering correction weight field. Based on the pixel neighborhood gradient of the water body data, the scattering correction weight field and the polarization component image are coupled and calculated to generate an initial compensation estimate. The initial compensation estimate is applied to the polarization component image, and the residual of the polarization component image before and after compensation is calculated. The residual is used as a supervision signal to train the pre-constructed compensation correction model. The polarization component images before and after compensation are input into the trained compensation and correction model, and the compensation and correction terms are output. Using the scattering correction weight field as the fusion coefficient, the initial compensation estimate and the compensation correction term are weighted and fused pixel by pixel, and the weighted fusion result is applied to the polarization multispectral image data to generate a compensation image; The compensated image is subjected to spectral correction and classification processing to generate plant classification results.

2. The method for accurate classification of plant species using image data processing algorithms according to claim 1, characterized in that: Acquiring water body data and polarization multispectral image data of the target object specifically includes: Acquire initial water body data and initial polarization multispectral image data of the target object; The initial polarization multispectral image data is compared frame by frame with the initial water body data corresponding to the time and space, and the initial polarization multispectral image data is corrected according to the frame-by-frame comparison results to generate polarization multispectral image data. The polarization multispectral image data is inverted, and the difference analysis of the polarization multispectral image data before and after inversion is performed. The difference analysis results are used as constraints to correct the initial water body data and generate water body data.

3. The method for accurate classification of plant species using image data processing algorithms according to claim 2, characterized in that: The process of mapping the water body data to the polarization component image pixel-by-pixel to generate a scattering correction weight field specifically includes: The water body data and the difference analysis results are input into a pre-built learner to generate an initial scattering correction weight for each pixel, which is then applied to the polarization component image and compared with the polarization component image. If the comparison result is greater than the preset discrimination threshold, the initial scattering correction weight is used as the scattering correction weight of the corresponding pixel, and then a scattering correction weight field is generated. Otherwise, the comparison results and the difference analysis results are fused together and input into the pre-built learner along with the water body data to iteratively generate initial scattering correction weights until the preset discrimination threshold is met, thereby generating a scattering correction weight field.

4. The method for accurate classification of plant species using image data processing algorithms according to claim 1, characterized in that: Training the pre-built compensation correction model using the residual as a supervision signal specifically includes: The scattering correction weight field and the residual are weighted and calculated to generate a weighted residual; The similarity between the residual and the residuals of pixels in its spatial neighborhood is calculated and coupled with the weighted residual. The coupling result is used as a supervision signal to train the pre-constructed compensation correction model.

5. The method for accurate classification of plant species using image data processing algorithms according to claim 1, characterized in that: The compensated image undergoes spectral correction and classification processing to generate plant classification results, specifically including: The compensated image and the polarization component image are spectrally fitted to generate a stable reflectance spectrum estimate; Cluster analysis is performed on the stable reflection spectrum estimates to generate a target spectrum time series set; The target spectral time series set and the compensated image are input into a pre-constructed fusion discriminator to output plant classification results.

6. The method for accurate classification of plant species using image data processing algorithms according to claim 5, characterized in that: Performing spectral fitting on the compensated image and the polarization component image to generate a stable reflectance spectrum estimate specifically includes: The physical consistency correction is performed on the compensated image based on the scattering correction weight field and the water body data to obtain the reference spectral value of each pixel; The compensated image and the polarization component image are spectrally fitted to generate an initial spectral estimate for each pixel, and the reference spectral value is subtracted from the estimate to generate a spectral residual. The temporal attention weight field is multiplied by the spectral residual, and the multiplication result is superimposed with the initial spectral estimate to generate a stable reflectance spectrum estimate for each pixel. The temporal attention weight field is obtained by calculating the sum of squares of the spectral residuals of each pixel at different time points.

7. A method for accurate plant species classification using image data processing algorithms according to claim 6, characterized in that: Cluster analysis is performed on the stable reflectance spectrum estimate to generate the target spectrum time series set, specifically including: Cluster analysis is performed on the stable reflectance spectrum estimates to form candidate spectral clusters; The temporal attention weight field is weighted with the stable reflectance spectrum estimate of each pixel in the candidate spectral cluster to generate the target spectral temporal set for each candidate spectral cluster.

8. A method for accurate classification of plant species using image data processing algorithms according to claim 7, characterized in that: After generating the target spectral time series set for each candidate spectral cluster, the process also includes calculating the motion consistency score for each candidate spectral cluster, specifically including: Based on the pixels contained in each candidate spectral cluster, the compensated image is subjected to target detection and segmentation processing to generate the pixel motion trajectory of each candidate spectral cluster. Based on the pixels contained in each candidate spectral cluster, extract the corresponding water body sub-data from the water body data; The pixel motion trajectory is compared with the corresponding water body sub-data to generate a motion consistency score for each candidate spectral cluster.

9. A method for accurate plant species classification using image data processing algorithms according to claim 8, characterized in that: The target spectral time series set and the compensated image are input into a pre-constructed fusion discriminator, and the output plant classification results specifically include: The target spectrum temporal set, the motion consistency score, the temporal attention weight field, and the compensation image are input into a pre-constructed fusion discriminator to output plant classification results. The pre-built fusion discriminator contains different data processing branches, each corresponding to a different discrimination strategy.

10. A precise plant species classification system utilizing image data processing algorithms, characterized in that, include: The data acquisition module is used to acquire water body data and polarization multispectral image data of the target object; The data processing module is used to perform pixel polarization decomposition on the polarization multispectral image data to generate polarization component images; The water body data is mapped pixel-by-pixel to the polarization component image to generate a scattering correction weight field. The initial compensation generation module is used to perform coupled calculations on the scattering correction weight field and the polarization component image based on the pixel neighborhood gradient of the water body data to generate an initial compensation estimate. The compensation correction term generation module is used to apply the initial compensation estimate to the polarization component image, calculate the difference between the polarization component images before and after compensation, and use the difference as a supervision signal to train the pre-constructed compensation correction model. The compensated polarization component image is input into the trained compensation and correction model, and the compensation and correction term is output. The compensation image generation module is used to use the scattering correction weight field as a fusion coefficient to perform weighted fusion of the initial compensation estimate and the compensation correction term pixel by pixel, and apply the weighted fusion result to the polarization multispectral image data to generate a compensation image. The plant classification module is used to perform spectral correction and classification processing on the compensated image to generate plant classification results.

Images