Water pollution trend prediction method based on three-dimensional fluorescence spectrum image features

Abstract

A water pollution trend prediction method based on three-dimensional fluorescence spectrum image features comprises the steps that firstly, the three-dimensional fluorescence spectrum image features are extracted through a Krawtchouk moment and an HOG descriptor, and the Euclidean distance between a current measurement point and a normal water quality cluster center is calculated and used for representing feature space similarity; meanwhile, analyzing the similarity of the position relationship through a binarization template of a fluorescence spectrum contour curve; and finally, carrying out weighted fusion on the two similarities to realize anomaly detection. For trend prediction, the pollution development trend is judged by calculating the included angle between the to-be-detected point and each pollution type cluster center relative to the normal water quality cluster center. According to the method, high-precision anomaly detection is realized by fusing the feature space similarity and the position relation similarity, and the pollution trend is judged by using the included angle. The method not only can capture the subtle change of the position and shape of the fluorescence peak of the three-dimensional fluorescence spectrum, but also can improve the recognition and prediction capabilities.

Description

Technical Field

[0001] This invention relates to a method for predicting water pollution trends based on three-dimensional fluorescence spectral image features. Background Technology

[0002] Water pollution is a global environmental problem that poses a serious threat to ecosystems and human health. With rapid industrialization and urbanization, the types and concentrations of pollutants in water bodies are becoming increasingly complex, and traditional monitoring methods are no longer sufficient to meet the needs of accurate early warning and trend prediction. Traditional methods for detecting water pollutants mainly rely on chemical analysis and biological detection methods. While these methods can provide relatively accurate results, they typically require long testing times and complex operating procedures.

[0003] In contrast, three-dimensional fluorescence spectroscopy exhibits significant advantages in water quality monitoring. By simultaneously scanning excitation and emission wavelengths, it can generate a unique "fluorescent fingerprint spectrum" that reflects the type and concentration of organic matter in the water. This makes three-dimensional fluorescence spectroscopy outstanding in terms of speed and high sensitivity. It eliminates the need for complex sample pretreatment, greatly shortening the detection cycle, reducing labor costs, and avoiding secondary pollution that may be introduced in traditional chemical analysis.

[0004] However, currently used methods, such as analysis based on fluorescence peak position or intensity, or extraction of fluorescence components using statistical methods like Parallel Factor Analysis (PARAFAC), struggle to fully capture subtle shape changes in the spectrum. While these methods simplify the complexity of the spectrum to some extent, they may result in insufficient ability to identify minute differences in complex organic matter composition and pollution changes in water bodies, thus affecting the accuracy of early warning and trend prediction. To overcome this deficiency, image processing techniques can be used to extract features from three-dimensional fluorescence spectral images. For example, Krawtchouk moments can describe the global geometric features of an image, while HOG descriptors excel at capturing local shape information. Combining the two can provide more comprehensive water quality image features. Subsequently, by calculating the angle between the test point and the cluster center of each pollution type relative to the cluster center of normal water quality, the pollution development trend can be determined. This method not only captures subtle changes but also improves the ability to identify and predict them.

[0005] Therefore, for the detection and early warning of water pollution anomalies, we need a water pollution trend prediction method based on three-dimensional fluorescence spectral image features to meet the needs of practical applications. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies in water quality anomaly detection and early warning, this invention provides a water pollution trend prediction method based on three-dimensional fluorescence spectral image features. This method achieves high-precision water quality anomaly detection by fusing feature spatial similarity and positional relationship similarity, and uses angle judgment to predict pollution trends.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] A water pollution trend prediction method based on three-dimensional fluorescence spectral image features is proposed. First, Krawtchouk moments and HOG descriptors are used to extract features from the three-dimensional fluorescence spectral image, and the Euclidean distance between the current measurement point and the normal water quality cluster center is calculated to characterize the feature space similarity. Simultaneously, the positional relationship similarity is analyzed by using the binarized template of the fluorescence spectral contour curve. Finally, the two similarities are weighted and fused to achieve anomaly detection. For trend prediction, the pollution development trend is determined by calculating the angle between the measurement point and the cluster center of each pollution type relative to the normal water quality cluster center.

[0009] The method includes the following steps:

[0010] Step 1: Calculate the similarity of image feature spaces

[0011] First, features are extracted from the grayscale image f(x,y) of the three-dimensional fluorescence spectrum using Krawtchouk moments to construct a feature vector, in which the element K... pq The calculation formula is as follows:

[0012]

[0013] Where p and q are the orders of the moments, p∈[0,P] max ],q∈[0,Q max f(x,y) is the gray value of the image at pixel (x,y); These are normalized Krawtchouk polynomials; p1 and p2 are the Krawtchouk polynomial parameters corresponding to the x and y axes, respectively; M and N are the length and width of the image, respectively; all calculated K... pq The values ​​are arranged in ascending order of p, then ascending order of q to form a new vector V = (K 00 ,K 01 ,…,K 0Qmax ,K 10 ,K 11 ,…,K PmaxQmax The dimension of V is U = (P) max +1)(Q max +1), re-represent the vector V as V=(k1,k2,…,k U ), where k uLet k1 be the u-th element in V, for example, k1 = K 00 k2=K 01 ,k Qmax+1 =K 0Qmax ,k U =K PmaxQmax And so on;

[0014] The grayscale image f(x,y) of the three-dimensional fluorescence spectrum was segmented using the HOG descriptor, and histograms of gradient orientations h of multiple local regions (blocks) were extracted. i After performing L2 normalization on each histogram, they are arranged in order to form the final feature vector V. HOG The calculation formula is as follows:

[0015]

[0016] V HOG =[v1,v2,...,v F (4)

[0017]

[0018] Among them, h i Represents the histogram count within the i-th region; ||h i ||2 represents the L2 norm of the histogram, used for normalization; region i Let represent the set of pixels in the i-th region of the image; δ(α) is the directional voting function, representing the contribution to one of the bins; bin in this method represents a specific directional range used to statistically analyze pixel gradient information within that range; ε represents a very small constant to avoid division by zero; v s It is a vector V HOG The s-th element in the array, s∈[1,F];

[0019] Then, vectors V and V HOG The first and last parts are concatenated to form a new vector T = (k1, k2, ..., k U ,v1,v2,…,v F Given a vector T with dimension U+F, re-represent the vector T as T = (t1, t2, ..., t...). U+F ), where t w Let t1 be the w-th element in T, for example, t1 = k1,t U+1 =v1,t U+F =v F And so on;

[0020] Feature extraction was performed on the three-dimensional fluorescence spectral data of normal water quality using Krawtchouk moments and HOG descriptors to obtain the cluster center feature vector C(c1,c2,c3,…,c…). U+F The dimension is U+F. For any measurement point, the Euclidean distance d between the feature vectors T and C is calculated to reflect the similarity of the feature space. The calculation formula is as follows:

[0021]

[0022] Among them, t j c is the j-th element of the feature vector of the current measurement point. j The j-th element of the eigenvector of the cluster center of normal water quality;

[0023] Step 2: Calculate the similarity of spectral positional relationships

[0024] The three-dimensional fluorescence spectra of the current measurement point and normal water quality are converted into contour curves and binarized to obtain binarized templates A(x,y) and B(x,y). Then, when A(x,y) and B(x,y) are multiplied, the position with a value of 1 is the intersection point.

[0025] R(x,y)=A(x,y)×B(x,y) (7)

[0026] Where R(x,y) satisfies:

[0027]

[0028] At each detected intersection point, draw the tangent lines to the two contour curves passing through that point, using the following formula:

[0029]

[0030] in, Let A represent the gradient vector of A at the point (x,y); Let B be the gradient vector at the point (x,y);

[0031] Since the tangent direction of the contour lines is perpendicular to the gradient, the following relationship is obtained:

[0032]

[0033] Among them, E A E represents the direction of the tangent line at point A(x,y). B This indicates the direction of the tangent line at point B(x,y). Solving for this, we get...

[0034]

[0035] Furthermore, calculate the angle θ between the two tangents at each intersection point. l Let l represent the included angle at the l-th intersection point, calculated using the following formula:

[0036]

[0037] Among them, the greater the shift or deformation of the fluorescence peak, the greater the corresponding θ. l The larger the value, the better θ l The value can effectively reflect subtle changes in the shape and position of the spectrum;

[0038] Finally, the acute angle θ l The absolute values ​​are summed and averaged to calculate the spectral position similarity Θ, as shown in the following formula:

[0039]

[0040] Where, θ l Let be the angle between the tangents at the l-th intersection point; m represents the number of angles.

[0041] Step 3: Calculate the overall similarity of water quality anomalies

[0042] The feature space similarity and the positional relationship similarity are weighted and fused to obtain the comprehensive similarity D, as shown in the following formula:

[0043] D=αd+βΘ (15)

[0044] Where α and β are weighting coefficients, satisfying α+β=1;

[0045] The calculated comprehensive similarity D is compared with the preset anomaly threshold D. th If D > D, then... th If the water quality at the current measurement point is determined to be in an abnormal state, an abnormality warning will be triggered immediately.

[0046] Step 4: Predict pollution trends

[0047] A feature database of e typical pollution types is obtained through prior knowledge, and the cluster center coordinates of the g-th typical pollution type are X. g =(x g1 ,x g2 ,x g3 ,…,x gL ), where L=U+F, and the coordinates of the normal water quality cluster center are C=(c1,c2,c3,…,c U+F For the three-dimensional fluorescence spectral characteristics of the test point, T = (t1, t2, ..., t...). U+F ), calculate T and each X g The angle ω between C and C g The calculation formula is as follows:

[0048]

[0049] ω g The smaller ω is, the higher the probability of the g-th type of pollution occurring. g When the minimum value is reached, the corresponding pollution type is the most probable development trend of the current measurement point.

[0050] This invention proposes a water pollution trend prediction method based on three-dimensional fluorescence spectral image features. This method achieves high-precision anomaly detection by fusing feature spatial similarity and positional similarity, and determines pollution trends by the included angle.

[0051] Preferably, in step 1, the value of ε is 1×10⁻⁶. -5 In step 3, the value of α is 0.372, the value of β is 0.628, and the anomaly threshold D is... th The value is 0.633.

[0052] The technical concept of this invention is as follows: A comprehensive anomaly detection method is introduced. This method extracts features from three-dimensional fluorescence spectral images using Krawtchouk moments and HOG descriptors, and calculates the Euclidean distance between the current measurement point and the normal water quality cluster center to characterize the feature space similarity. Simultaneously, the positional relationship similarity is analyzed using a binary template of the fluorescence contour curve. Finally, the two similarities are weighted and fused to achieve comprehensive anomaly detection. A pollution evolution trend prediction method is also proposed, which determines the pollution development trend by calculating the angle between the measurement point and the cluster center of each pollution type relative to the normal water quality cluster center.

[0053] The beneficial effects of this invention are:

[0054] 1. By fusing feature space similarity and positional relationship similarity, the accuracy and universality of water quality anomaly detection are improved, enabling effective detection of various water quality anomaly features.

[0055] 2. An adaptive weighting mechanism is used to optimize the overall similarity, ensuring the ability to comprehensively identify complex water quality anomalies. Attached Figure Description

[0056] Figure 1 This is a flowchart of the method of the present invention.

[0057] Figure 2 The feature space is constructed using Krawtchouk moments and HOG descriptors.

[0058] Figure 3 It is a binary template image generated using fluorescence contour plots.

[0059] Figure 4This is a schematic diagram illustrating the calculation of positional similarity.

[0060] Figure 5 This is a schematic diagram for calculating the included angle of water quality abnormality deviation. Figure 6 This is a diagram illustrating the prediction of pollution trends. Detailed Implementation

[0061] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0062] Reference Figures 1-5 This paper presents a water pollution trend prediction method based on three-dimensional fluorescence spectral image features. The method first constructs a feature space using Krawtchouk moments and HOG descriptors, calculating the Euclidean distance between the current measurement point and the normal water quality cluster center. Simultaneously, it analyzes the positional similarity using a binarized template of the fluorescence contour curves. Finally, it weights and fuses the two similarities to achieve comprehensive anomaly detection. For trend prediction, it determines the pollution development trend by calculating the angle between the measurement point and the cluster center of each pollution type relative to the normal water quality cluster center, referencing... Figure 1 .

[0063] The water pollution trend prediction method based on three-dimensional fluorescence spectral image features includes the following steps:

[0064] Step 1: Calculate the similarity of image feature spaces

[0065] First, features are extracted from the grayscale image f(x,y) of the three-dimensional fluorescence spectrum using Krawtchouk moments to construct a feature vector. The element K in this vector... pq The calculation formula is as follows:

[0066]

[0067] Where p and q are the orders of the moments, p∈[0,P] max ],q∈[0,Q max f(x,y) is the gray value of the image at pixel (x,y); These are normalized Krawtchouk polynomials; p1 and p2 are the Krawtchouk polynomial parameters corresponding to the x and y axes, respectively; M and N are the length and width of the image, respectively. All calculated K... pq The values ​​are arranged in ascending order of p, then ascending order of q to form a new vector V = (K 00 ,K 01 ,…,K 0Qmax ,K 10 ,K 11 ,…,K PmaxQmax The dimension of V is U = (P) max +1)(Q max+1), re-represent the vector V as V=(k1,k2,…k U ), where k u Let k1 be the u-th element in V, for example, k1 = K 00 k2=K 01 ,k Qmax+1 =K 0Qmax ,k U =K PmaxQmax And so on;

[0068] The grayscale image f(x,y) of the three-dimensional fluorescence spectrum was segmented using the HOG descriptor, and histograms of gradient orientations h of multiple local regions (blocks) were extracted. i After performing L2 normalization on each histogram, they are arranged in order to form the final feature vector V. HOG ,refer to Figure 2 The calculation formula is as follows:

[0069]

[0070] V HOG =[v1,v2,...,v F (20)

[0071]

[0072] Among them, h i Represents the histogram count within the i-th region; ||h i ||2 represents the L2 norm of the histogram, used for normalization; region i Let represent the set of pixels in the i-th region of the image; δ(α) is the directional voting function, representing the contribution to one of the bins; bin in this method represents a specific directional range used to statistically analyze pixel gradient information within that range; ε represents a very small constant to avoid division by zero; v s It is a vector V HOG The s-th element in the array, s∈[1,F];

[0073] Then, vectors V and V HOG The first and last parts are concatenated to form a new vector T = (k1, k2, ..., k U ,v1,v2,…,v F Given a vector T with dimension U+F, re-represent the vector T as T = (t1, t2, ..., t...). U+F ). Where t w Let t1 be the w-th element in T, for example, t1 = k1,t U+1 =v1,t U+F =v F And so on;

[0074] Feature extraction was performed on the three-dimensional fluorescence spectral data of normal water quality using Krawtchouk moments and HOG descriptors to obtain the cluster center feature vector C(c1,c2,c3,…,c…). U+F The dimension is U+F. For any measurement point, the Euclidean distance d between the feature vectors T and C is calculated to reflect the similarity of the feature space. (Refer to...) Figure 3 The calculation formula is as follows:

[0075]

[0076] Among them, t j c is the j-th element of the feature vector of the current measurement point. j The j-th element of the eigenvector of the cluster center of normal water quality;

[0077] Step 2: Calculate the similarity of spectral positional relationships

[0078] The three-dimensional fluorescence spectra of the current measurement point and normal water quality are converted into contour plots and binarized to obtain binarized templates A(x,y) and B(x,y). (Refer to...) Figure 4 Then, when A(x,y) and B(x,y) are multiplied, the position where the value is 1 is the intersection point.

[0079] R(x,y)=A(x,y)×B(x,y) (23)

[0080] Where R(x,y) satisfies:

[0081]

[0082] At each detected intersection point, draw the tangent lines to the two contour curves passing through that point, using the following formula:

[0083]

[0084] in, Let A represent the gradient vector of A at the point (x,y); Let B be the gradient vector at the point (x,y);

[0085] Since the tangent direction of the contour lines is perpendicular to the gradient, the following relationship is obtained:

[0086]

[0087] Among them, E A E represents the direction of the tangent line at point A(x,y). B Indicate the direction of the tangent line at point B(x,y), and solve for...

[0088]

[0089] Furthermore, calculate the angle θ between the two tangents at each intersection point. l , l represents the included angle at the l-th intersection point, refer to Figure 5 The calculation formula is as follows:

[0090]

[0091] Among them, the greater the shift or deformation of the fluorescence peak, the greater the corresponding θ. l The larger the value, the better θ l The value can effectively reflect subtle changes in the shape and position of the spectrum;

[0092] Finally, the acute angle θ l The absolute values ​​are summed and averaged to calculate the spectral position similarity Θ. The calculation formula is as follows:

[0093]

[0094] Where, θ l Let be the angle between the tangents at the l-th intersection point; m represents the number of angles.

[0095] Step 3: Calculate the overall similarity of water quality anomalies

[0096] The feature space similarity and the positional relationship similarity are weighted and fused to obtain the comprehensive similarity D, as shown in the following formula:

[0097] D=αd+βΘ (31)

[0098] Where α and β are weighting coefficients, satisfying α+β=1;

[0099] The calculated comprehensive similarity D is compared with the preset anomaly threshold D. th Compare them. If D > D th If the water quality at the current measurement point is determined to be in an abnormal state, an abnormality warning will be triggered immediately.

[0100] Step 4: Predict pollution trends

[0101] A feature database of e typical pollution types is obtained through prior knowledge, and the cluster center coordinates of the g-th typical pollution type are X. g =(x g1 ,x g2 ,x g3 ,…,x gL ), where L=U+F, and the coordinates of the normal water quality cluster center are C=(c1,c2,c3,…,c U+F For the three-dimensional fluorescence spectral characteristics of the test point, T = (t1, t2, ..., t...). U+F), calculate T and each X g The angle ω between C and C g ,refer to Figure 6 The calculation formula is as follows:

[0102]

[0103] ω g The smaller ω is, the higher the probability of the g-th type of pollution occurring. g When the minimum value is reached, the corresponding pollution type is the most probable development trend of the current measurement point.

[0104] Based on this, this invention introduces an adaptive weighted optimization mechanism to further improve the accuracy and robustness of water quality anomaly detection and trend prediction. By evaluating the performance of feature space similarity and location relationship similarity under different weight configurations, the weighting coefficients α and β are dynamically adjusted using an optimization algorithm to identify the optimal weight combination. After obtaining the optimal weight configuration, a water pollution trend prediction method based on three-dimensional fluorescence spectral image features is formed by combining historical data and real-time monitoring information. This method can not only more accurately identify water quality anomalies but also effectively predict pollution evolution trends, providing more reliable decision support for water environment management.

[0105] This embodiment takes detecting the type of pollution in a river surrounding a pharmaceutical factory as an example. The specific implementation steps are as follows:

[0106] Step 1: Collect samples and obtain their three-dimensional fluorescence spectra. Use Krawtchouk moments and HOG descriptors to extract the feature T of the grayscale image of the spectrum. Calculate the Euclidean distance d between T and the cluster center C of normal water quality to characterize the feature space similarity.

[0107] Step 2: Binarize the fluorescence spectra of the current measurement point and the normal water quality, and then multiply them; at each detected intersection point, draw the tangents of the two contour curves passing through that point, and calculate the angle θ between them. l The spectral position similarity Θ is calculated by summing the absolute values ​​of all included angles and taking the average.

[0108] Step 3: Perform a weighted fusion of the feature space similarity obtained in Step 1 and the positional relationship similarity obtained in Step 2 to obtain a comprehensive similarity D. The calculated D = 0.821. Then, compare D with the preset anomaly threshold D0. th Comparing, D>D th Therefore, the river is polluted;

[0109] Step 4: Predict pollution trends

[0110] Calculations show that, under T and various pollution types X... gThe angle ω between C and C g In the figure, ω3 = 2.784° is the minimum value. According to the database query, when g = 3, the pollution type is antibiotic pollution. Therefore, it can be determined that there is antibiotic pollution in the river.

[0111] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for predicting water pollution trends based on three-dimensional fluorescence spectral image features, characterized in that, First, Krawtchouk moments and HOG descriptors are used to extract features from three-dimensional fluorescence spectral images. The Euclidean distance between the current measurement point and the cluster center of normal water quality is calculated to characterize the feature space similarity. Meanwhile, the similarity of positional relationships is analyzed by using the binary template of the contour curves of fluorescence spectra; finally, the two similarities are weighted and fused to achieve anomaly detection; for trend prediction, the pollution development trend is judged by calculating the angle between the test point and the cluster center of each pollution type relative to the cluster center of normal water quality.

2. The water pollution trend prediction method based on three-dimensional fluorescence spectral image features as described in claim 1, characterized in that, The method includes the following steps: Step 1: Calculate the similarity of image feature spaces First, features are extracted from the grayscale image f(x,y) of the three-dimensional fluorescence spectrum using Krawtchouk moments to construct a feature vector, in which the element K... pq The calculation formula is as follows: Where p and q are the orders of the moments, p∈[0,P] max ],q∈[0,Q max f(x,y) is the gray value of the image at pixel (x,y); These are normalized Krawtchouk polynomials; p1 and p2 are the Krawtchouk polynomial parameters corresponding to the x and y axes, respectively; M and N are the length and width of the image, respectively; all calculated K... pq The values ​​are arranged in ascending order of p, then ascending order of q to form a new vector V = (K 00 ,K 01 ,…,K 0Qmax ,K 10 ,K 11 ,…,K PmaxQmax The dimension of V is U = (P) max +1)(Q max +1), re-represent the vector V as V=(k1,k2,…,k U ), where k u Let k1 be the u-th element in V, for example, k1 = K 00 k2=K 01 ,k Qmax+1 =K 0Qmax ,k U =K PmaxQmax And so on; The grayscale image f(x,y) of the three-dimensional fluorescence spectrum was segmented using the HOG descriptor, and histograms of gradient orientations h of multiple local regions (blocks) were extracted. i After performing L2 normalization on each histogram, they are arranged in order to form the final feature vector V. HOG The calculation formula is as follows: V HOG =[v1,v2,...,v F ] (4) Among them, h i Represents the histogram count within the i-th region; ||h i ||2 represents the L2 norm of the histogram, used for normalization; region i Let represent the set of pixels in the i-th region of the image; δ(α) is the directional voting function, representing the contribution to one of the bins; bin in this method represents a specific directional range used to statistically analyze pixel gradient information within that range; ε represents a very small constant to avoid division by zero; v s It is a vector V HOG The s-th element in the array, s∈[1,F]; Then, vectors V and V HOG The first and last parts are concatenated to form a new vector T = (k1, k2, ..., k U ,v1,v2,…,v F Given a vector T with dimension U+F, re-represent the vector T as T = (t1, t2, ..., t...). U+F ), where t w Let t1 be the w-th element in T, for example, t1 = k1,t U+1 =v1,t U+F =v F And so on; Feature extraction was performed on the three-dimensional fluorescence spectral data of normal water quality using Krawtchouk moments and HOG descriptors to obtain the cluster center feature vector C(c1,c2,c3,…,c…). U+F The dimension is U+F. For any measurement point, the Euclidean distance d between the feature vectors T and C is calculated to reflect the similarity of the feature space. The calculation formula is as follows: Among them, t j c is the j-th element of the feature vector of the current measurement point. j The j-th element of the eigenvector of the cluster center of normal water quality; Step 2: Calculate the similarity of spectral positional relationships The three-dimensional fluorescence spectra of the current measurement point and normal water quality are converted into contour curves and binarized to obtain binarized templates A(x,y) and B(x,y). Then, when A(x,y) and B(x,y) are multiplied, the position with a value of 1 is the intersection point. R(x,y)=A(x,y)×B(x,y) (7) Where R(x,y) satisfies: At each detected intersection point, draw the tangent lines to the two contour curves passing through that point, using the following formula: in, Let A represent the gradient vector of A at the point (x,y); Let B be the gradient vector at the point (x,y); Since the tangent direction of the contour lines is perpendicular to the gradient, the following relationship is obtained: Among them, E A E represents the direction of the tangent line at point A(x,y). B Indicate the direction of the tangent line at point B(x,y), and solve for... Furthermore, calculate the angle θ between the two tangents at each intersection point. l Let l represent the included angle at the l-th intersection point, calculated using the following formula: Among them, the greater the shift or deformation of the fluorescence peak, the greater the corresponding θ. l The larger the value, the better θ l The value can effectively reflect subtle changes in the shape and position of the spectrum; Finally, the acute angle θ l The absolute values ​​are summed and averaged to calculate the spectral position similarity Θ, as shown in the following formula: Where, θ l Let be the angle between the tangents at the l-th intersection point; m represents the number of angles. Step 3: Calculate the overall similarity of water quality anomalies The feature space similarity and the positional relationship similarity are weighted and fused to obtain the comprehensive similarity D, as shown in the following formula: D=αd+βΘ (15) Where α and β are weighting coefficients, satisfying α+β=1; The calculated comprehensive similarity D is compared with the preset anomaly threshold D. th If D > D, then... th If the water quality at the current measurement point is determined to be in an abnormal state, an abnormality warning will be triggered immediately. Step 4: Predict pollution trends A feature database of e typical pollution types is obtained through prior knowledge, and the cluster center coordinates of the g-th typical pollution type are X. g =(x g1 ,x g2 ,x g3 ,…,x gL ), where L=U+F, and the coordinates of the normal water quality cluster center are C=(c1,c2,c3,…,c U+F For the three-dimensional fluorescence spectral characteristics of the test point, T = (t1, t2, ..., t... U+F ), calculate T and each X g The angle ω between C and C g The calculation formula is as follows: When ω g When the minimum value is reached, the corresponding pollution type is the most probable development trend of the current measurement point.

3. The water pollution trend prediction method based on three-dimensional fluorescence spectral image features as described in claim 2, characterized in that, In step 1, the value of ε is 1×10. -5 In step 3, the value of α is 0.372, the value of β is 0.628, and the anomaly threshold D is... th The value is 0.633.

Images