Urban black and odorous water body water quality judgment method based on hyperspectral remote sensing data deep learning model

Abstract

The invention discloses an urban black and odorous water quality judgment method based on a hyperspectral remote sensing data deep learning model, and particularly relates to the technical field of water quality judgment. Water quality monitoring data of a plurality of sampling points and corresponding hyperspectral remote sensing images are obtained, a water body dynamic change sensitivity index of each sampling point is calculated based on historical water quality change conditions, an end member extraction algorithm is adopted to analyze spectrum mixing conditions of remote sensing pixels, and the spectral purity confidence coefficient of the remote sensing pixels is calculated; the method comprises the following steps of: acquiring a plurality of samples, generating a credibility grade of each sample, removing or reducing the weight of low-credibility samples, introducing a time sequence remote sensing image for the samples in a high water body dynamic change sensitivity index region, and reconstructing remote sensing features through interpolation to recover a spectrum state at a real sampling moment, so that the reliability of remote sensing-actual measurement fusion data is remarkably improved; and the judgment capability of the model in a water quality rapid change or interference complex region is enhanced, and the remote sensing intelligent identification of the urban black and odorous water body with higher precision and stronger robustness is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of water quality determination, and in particular to a method for determining the water quality of urban black and odorous water bodies based on a deep learning model of hyperspectral remote sensing data. Background Art

[0002] Determination of urban black and odorous water quality involves testing and evaluating water quality indicators in urban water bodies (such as rivers, lakes, and ditches) to determine whether they are noticeably black or have a foul odor. This determination typically relies on a combination of sensory evaluation (such as water color and odor) and physical and chemical indicators (such as dissolved oxygen, ammonia nitrogen, and COD). The goal is to identify the severity of urban water pollution and provide a basis for managing and improving the aquatic ecosystem.

[0003] The existing technology has the following shortcomings:

[0004] When aligning the spectral signatures of sampling points in remote sensing imagery with their field-based water quality data, a water quality mismatch can occur due to spatiotemporal asynchrony. This problem stems from inconsistencies in the acquisition time or space between remote sensing imagery and measured water quality data. This results in the remote sensing signatures and measured indicators at the same location not reflecting the same water quality state. This is particularly pronounced in environments with rapidly changing water bodies or strong spatial heterogeneity. Furthermore, this mismatch introduces label noise, misleading the model into learning incorrect feature relationships, reducing classification or regression accuracy, and ultimately affecting the model's generalization and stability in practical applications. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for determining the water quality of urban black and odorous water bodies based on a deep learning model of hyperspectral remote sensing data to address the shortcomings of the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for determining the water quality of urban black and odorous water bodies based on a deep learning model of hyperspectral remote sensing data, comprising:

[0007] Obtain water quality monitoring data and remote sensing images from several sampling points;

[0008] Based on water quality monitoring data, the frequency and magnitude of water quality changes in the area where the sampling points are located are analyzed, and the sensitivity index of water body dynamic changes is calculated;

[0009] Use the endmember extraction algorithm to obtain the spectral mixing degree of the remote sensing pixel corresponding to the sampling point, determine its interference degree, and calculate the confidence level of the spectral purity of the remote sensing pixel;

[0010] The water body dynamic change sensitivity index and remote sensing pixel spectral purity confidence level corresponding to each sampling point are comprehensively scored to generate the sample credibility level, and low-credibility samples are eliminated;

[0011] For samples in areas where the sensitivity index of water body dynamic changes is greater than the set threshold, time series images are introduced for interpolation to reconstruct the remote sensing characteristics of the sampling time points and perform confidence correction on the labels;

[0012] The water body dynamic change sensitivity index and remote sensing pixel spectral purity confidence of the sample are used as weighting factors of the deep learning model loss function to train the deep learning model;

[0013] Use the trained deep learning model to determine the black and odorous water quality of urban water bodies.

[0014] Preferably, obtaining water quality monitoring data includes: recording the sampling time during the field sampling process, using differential GPS to record the geographic coordinates of the sampling points, and the positioning error is less than 1 meter; synchronously collecting hydrological environment information, including water flow velocity, distribution of surrounding sewage outlets, rainfall in the 24 hours before sampling, water color and turbidity.

[0015] Preferably, the method for obtaining the water body dynamic change sensitivity index WVSI is: obtaining multiple water quality measured values ​​of the sampling point or its surrounding adjacent pixels within a continuous time period; including COD, NH3-N and DO; calculating the change frequency of the water quality monitoring data: F is the frequency of change, N var is the number of samples with significant changes, N total is the total sample number;

[0016] Calculate the change range of water quality monitoring data: A is the average change amplitude, Q i is the water quality index value of the i-th sampling in the time series, and N is the number of sequence samples; the change frequency and change amplitude of the water quality monitoring data are calculated for different indicators respectively, and the water body dynamic change sensitivity index WVSI is obtained by normalizing them and performing weighted summation.

[0017] Preferably, an endmember extraction algorithm is used to obtain the spectral mixing degree of the remote sensing pixel corresponding to the sampling point, determine the degree of interference, and calculate the confidence level of the spectral purity of the remote sensing pixel, specifically:

[0018] Use the water extraction algorithm to extract the water area in the remote sensing image; locate each sampling point to the corresponding remote sensing pixel and its surrounding pixel set for analysis; the goal is to identify the purest spectral pixel as the end member; use the spectral angle or linear mixture model to calculate the similarity between the pixel corresponding to the current sampling point and the purest water end member, the expression is: Where, is the reflectivity vector of the sampling point pixel, is the spectral vector of the water end member, θ is the spectral angle, and the unit is radian; the spectral purity confidence SPC of the remote sensing pixel is calculated by reverse normalization according to the spectral angle: SPC∈[0,1],θ max It is the maximum spectral angle of the water pixel in the dataset and is used for normalization.

[0019] Preferably, WVSI and SPC are normalized and the sample credibility score is calculated, and the expression is: SCL = λ·WVSI′+(1-λ)·SPC′; where SPC′ is the spectral purity confidence of the remote sensing pixel after normalization; WVSI′ is the water body dynamic change sensitivity index after normalization; SCL is the sample credibility level, ranging from [0,1]; λ is the weight coefficient.

[0020] Preferably, samples are divided into different levels according to the comprehensive score SCL:

[0021] High confidence level: SCL ≥ 0.8, the sample is stable and the observation is pure, and can be used directly for training;

[0022] Medium credibility level: 0.6≤SCL<0.8, the sample is credible but has slight interference and can be used with a lower priority;

[0023] Low confidence level: SCL < 0.6, the samples are potentially mismatched or mixed and should be eliminated or used for uncertainty modeling.

[0024] Preferably, for samples in areas where the water body dynamic change sensitivity index is greater than a set threshold, time series images are introduced for interpolation to reconstruct the remote sensing characteristics of the sampling time points and perform confidence correction on the labels:

[0025] Traverse all sampling points and identify their corresponding WVSI values;

[0026] The samples whose WVSI exceeds the set threshold are classified into the sensitive area set;

[0027] The reliability of the synchronization between the remote sensing features and the measured labels of the samples in the collection is low and needs to be processed;

[0028] Collect remote sensing images of the same area within the time range before and after the sampling time point;

[0029] For each high WVSI sample, extract the spectral vector of its corresponding pixel at multiple time points; generate a time series curve for each band: Among them, R b (t) is the reflectivity of band b at time t; this operation is performed for all bands b∈[1,N]; T-1, T0, T+1 are three different time points;

[0030] For the sampling time point t sampleThe spectral reflectance at is estimated: Reconstruct the entire remote sensing feature vector: N is the total number of bands; replace the original remote sensing features Used for sample matching;

[0031] Perform difference analysis between original remote sensing features and reconstructed remote sensing features: δ is the consistency deviation; ∈ indicates preventing division by zero and is set to the minimum value;

[0032] According to the degree of deviation, the confidence of the sample label is adjusted, and the reconstructed remote sensing features are used. As model input, and use the corrected labels or confidences to participate in training.

[0033] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0034] 1. This invention introduces two novel indicators: the Water Variation Sensitivity Index (WVSI) and the Spectral Purity Confidence of Remotely Sensed Pixels (SPC), to achieve dual assessment and fine-grained screening of the quality of remotely sensed and measured water quality samples, effectively resolving the water quality mismatch problem caused by spatiotemporal asynchrony in existing technologies. Compared to traditional methods that simply rely on remote sensing data at a single time point, this invention considers water quality dynamics and observation purity during the sample construction phase, and significantly improves the consistency and effectiveness of training data through credibility level classification, sample elimination, or downgrading.

[0035] 2. This method uses time-series remote sensing imagery in high WVSI areas for interpolation and reconstruction, dynamically repairs remote sensing features, and applies confidence correction to labels, further reducing the interference of label noise on model training. By incorporating WVSI and SPC into the loss function of the deep learning model as weighting factors, the model is guided to focus on learning high-confidence samples, improving the accuracy, robustness, and generalization of the model's determination of black and odorous water quality in complex urban water environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0037] Figure 1 This is a mind map of the method of the present invention. DETAILED DESCRIPTION

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0039] For examples, see Figure 1 As shown, the method for determining the water quality of urban black and odorous water bodies based on a deep learning model of hyperspectral remote sensing data described in this embodiment includes:

[0040] Obtain water quality monitoring data and remote sensing images from several sampling points;

[0041] Based on water quality monitoring data, the frequency and magnitude of water quality changes in the area where the sampling points are located are analyzed, and the sensitivity index of water body dynamic changes is calculated;

[0042] Use the endmember extraction algorithm to obtain the spectral mixing degree of the remote sensing pixel corresponding to the sampling point, determine its interference degree, and calculate the confidence level of the spectral purity of the remote sensing pixel;

[0043] The water body dynamic change sensitivity index and remote sensing pixel spectral purity confidence level corresponding to each sampling point are comprehensively scored to generate the sample credibility level, and low-credibility samples are eliminated;

[0044] For samples in areas where the sensitivity index of water body dynamic changes is greater than the set threshold, time series images are introduced for interpolation to reconstruct the remote sensing characteristics of the sampling time points and perform confidence correction on the labels;

[0045] The water body dynamic change sensitivity index and remote sensing pixel spectral purity confidence of the sample are used as weighting factors of the deep learning model loss function to train the deep learning model;

[0046] Use the trained deep learning model to determine the black and odorous water quality of urban water bodies.

[0047] During the field water quality sampling process, accurate environmental information must be recorded for each sampling point to enhance the accuracy of subsequent remote sensing data matching. Specifically, this includes:

[0048] Timestamp recording: Use high-precision time synchronization equipment (such as Beidou / GPS timing terminal) to record the exact time of water quality sampling. The accuracy should reach the minute or even second level to match the corresponding remote sensing imaging moment.

[0049] Spatial location recording: Use a GPS module or differential GPS (RTK) system to measure the geographic coordinates (longitude and latitude) of the sampling points. The coordinate accuracy is recommended to be better than ±1 meter to ensure accurate correspondence with remote sensing pixels.

[0050] Hydrological environment information collection:

[0051] Water flow velocity: Measure water flow velocity using a portable current meter or float range finding method to assess pollutant transport rates;

[0052] Whether it is near the sewage outlet: record on site whether the sampling point is located near the urban sewage outlet, rainwater drainage channel or industrial pipeline;

[0053] Weather and rainfall conditions: record the rainfall conditions 24 hours before sampling and on the day of sampling (can be connected to the meteorological API or local rain gauge data) to determine the dilution or disturbance of water bodies;

[0054] Other information: such as water depth, turbidity perception, whether there are obvious floating objects, etc.

[0055] These data will be used as auxiliary input variables for subsequent water body change sensitivity assessment or sample weighting processing.

[0056] Remote sensing image selection:

[0057] Select hyperspectral remote sensing images that are as synchronized as possible with the field sampling time or with a time difference of less than 3 hours;

[0058] The data source can be a hyperspectral sensor carried by a UAV, an airborne imaging spectrometer, or a satellite platform (such as GF-5, PRISMA, Hyperion, etc.);

[0059] The images should have high spectral resolution (better than 10 nm) and appropriate spatial resolution (better than 10 m) to meet the needs of small-scale water body analysis.

[0060] Radiometric correction: converting raw DN values ​​into radiance or reflectance, using the FLAASH, QUAC or Empirical Line methods;

[0061] Atmospheric correction: eliminates interference from atmospheric scattering, water vapor, aerosols, etc. on remote sensing signals, and improves the accuracy of water body reflectivity;

[0062] Geometric correction and registration: The remote sensing image is geometrically corrected to match the actual geographic coordinate system and accurately registered with the GPS sampling points. The error should be controlled within 1 pixel.

[0063] Water body extraction and mask processing: Apply water body indices such as NDWI or MNDWI to eliminate non-water body areas and retain only valid water body pixels;

[0064] Cloud and shadow removal: Identify and remove abnormal areas such as clouds, shadows, and vegetation obstructions to ensure that the area corresponding to the sampling point is observable.

[0065] Check whether the target sampling point is in the observation blind area or edge pixel area; evaluate the image clarity, band integrity and signal-to-noise ratio of the target area to decide whether to include it in the sample construction.

[0066] Based on the water quality monitoring data, the frequency and amplitude of water quality changes in the area where the sampling points are located are analyzed, and the sensitivity index of water body dynamic changes is calculated, specifically:

[0067] The Water Variation Sensitivity Index (WVSI) quantifies the intensity and frequency of water quality fluctuations at a sampling point or region over a specific timescale, assessing its sensitivity to pollution disturbances and hydrological impacts. A higher WVSI indicates more dramatic water quality fluctuations in the region, and a greater likelihood of mismatches between remotely sensed and measured water quality.

[0068] The acquisition method is: obtain multiple water quality measurements at the sampling point or its surrounding adjacent pixels within a continuous time period; the time scale is recommended to cover nearly 1 day to 7 days, and the higher the temporal resolution, the better (such as hourly level); select one or more representative indicators (such as COD, NH3-N, DO).

[0069] Calculate the frequency of changes in water quality monitoring data: F is the frequency of change (unit: dimensionless, 0~1), N var is the number of samples with significant changes (changes exceeding the threshold), N total is the total number of samples, and Δ is the set change threshold, for example, the intraday change of COD is greater than ±15%.

[0070] Calculate the change range of water quality monitoring data: A is the average change (the unit is consistent with the water quality index, such as mg / L), Q i is the water quality index value sampled at the i-th time in the time series, and N is the number of sequence samples; the change frequency and change amplitude of water quality monitoring data of different indicators (such as COD, NH3-N, DO) are calculated separately, and the water body dynamic change sensitivity index is obtained by normalizing them and performing weighted summation.

[0071] A higher WVSI indicates that water quality in a region experiences high frequency or amplitude fluctuations over short periods of time, potentially due to factors such as rainfall, sewage discharge, and flow rate, leading to rapid changes in water quality. This instability can lead to a timing mismatch between measured water quality data and remote sensing observations, reducing the consistency and credibility of remote sensing-measured samples.

[0072] Conversely, a smaller WVSI value indicates relatively stable water quality in the area, with lower fluctuations and frequency, and the remote sensing imagery is more likely to accurately reflect the water quality at the time of sampling. Therefore, the smaller the WVSI, the higher the sample's credibility, making it more suitable as training data for deep learning models.

[0073] The endmember extraction algorithm is used to obtain the spectral mixing degree of the remote sensing pixel corresponding to the sampling point, determine its interference degree, and calculate the confidence level of the spectral purity of the remote sensing pixel. Specifically:

[0074] Use water body extraction algorithms (such as NDWI and MNDWI) to extract water body areas in remote sensing images; locate each sampling point to the corresponding remote sensing pixel and its surrounding pixel set (such as a 3×3 or 5×5 window) for analysis.

[0075] Use endmember extraction algorithms, such as PPI (Pixel Purity Index); N-FINDR; VCA (Vertex Component Analysis);

[0076] The goal is to identify the purest spectral pixels as endmembers, representing typical substances such as water, vegetation, soil, and buildings.

[0077] The spectral angle or linear mixed model is used to calculate the similarity between the pixel corresponding to the current sampling point and the end member of the purest water body. The expression is: Where, is the reflectance vector (multi-band) of the sampling point pixel, is the water endmember spectral vector, θ is the spectral angle in radians, the smaller it is, the closer it is to the water endmember;

[0078] Calculate the spectral purity confidence SPC of remote sensing pixels and obtain it by reverse normalization according to the spectral angle: SPC∈[0,1], the closer it is to 1, the purer the spectrum; θ max It is the maximum spectral angle of the water pixel in the dataset and is used for normalization.

[0079] A higher spectral purity confidence (SPC) value for a remote sensing pixel indicates that the pixel's spectral signature is closer to that of pure water, and is less likely to be affected by non-water objects such as shore, vegetation, floating debris, or shadows. A high SPC value indicates that the remote sensing information reflected by the pixel is more reliable and accurately represents the spectral characteristics of the water itself, thereby improving the overall credibility of the sample.

[0080] Conversely, when the SPC value is small, it indicates that the pixel has a high degree of spectral mixing or interference factors, and the remote sensing characteristics may contain non-water body information, resulting in inaccurate water quality judgment of the sample by the model. Therefore, the smaller the SPC value, the lower the sample credibility, and it is recommended to downgrade or remove it during training.

[0081] In order to improve the overall quality of remote sensing-measured water quality samples and reduce errors caused by remote sensing mismatch and mixed pixels, the water body dynamic change sensitivity index WVSI and the remote sensing pixel spectral purity confidence SPC corresponding to each sampling point are comprehensively evaluated to generate the sample credibility level SCL, and sample elimination or weighted processing is performed accordingly.

[0082] WVSI and SPC were normalized to ensure that the two indices were on the same scale [0, 1] using the following method: WVSI′ = 1-WVSI; the larger the original WVSI value, the stronger the instability, and it needs to be negated to make it larger and more reliable. WVSI′ is the sensitivity index of water body dynamic changes after normalization; SPC′ = SPC itself is a confidence-type indicator and does not need to be transformed.

[0083] The sample confidence score is calculated using the expression: SCL = λ·WVSI′ + (1-λ)·SPC′; where SPC′ is the confidence level of the spectral purity of the normalized remote sensing pixel; SCL is the sample confidence level, ranging from [0,1]; and λ is the weight coefficient (the recommended value is 0.5-0.7, with SPC representing the observation quality). For example, if λ = 0.6, water quality stability accounts for 60% and spectral purity accounts for 40%.

[0084] According to the comprehensive score SCL, the samples are divided into different levels:

[0085] High confidence level: SCL ≥ 0.8, the sample is stable and the observation is pure, and can be used directly for training;

[0086] Medium credibility level: 0.6≤SCL<0.8, the sample is credible but has slight interference and can be used with a lower priority;

[0087] Low confidence level: SCL < 0.6, the samples are potentially mismatched or mixed and should be eliminated or used for uncertainty modeling.

[0088] High-confidence samples are retained intact and used directly for model training. Medium-confidence samples can be downgraded, such as by setting a weight of 0.5 or introducing a noise suppression mechanism during training. Low-confidence samples are discarded by default and do not participate in supervised training, but can be used for data augmentation simulations, uncertainty training (such as MC-Dropout or fuzzy labeling), and model robustness testing. Sample distribution maps or heat maps are generated, with different SCL regions marked, to aid decision-making, such as sampling point selection and optimization and model error analysis.

[0089] For samples in areas where the sensitivity index of water body dynamic changes is greater than the set threshold, time series images are introduced for interpolation, the remote sensing characteristics of the sampling time points are reconstructed, and the labels are confidence corrected.

[0090] Traverse all sampling points and identify their corresponding WVSI values;

[0091] Samples whose WVSI exceeds the set threshold (e.g., 0.7) are classified into the sensitive area set;

[0092] The synchronization between the remote sensing features and the measured labels of the samples in the collection is of low reliability and needs to be processed.

[0093] Collect remote sensing images of the same area within a time range before and after the sampling time point (e.g., ±3 days or ±12 hours);

[0094] Images are required to have the same band settings, the same sensor source, or unified atmospheric correction processing to ensure temporal consistency; and contain image data from at least three or more time nodes (T-1, T0, T+1).

[0095] For each high WVSI sample, extract the spectral vector (multi-band) of its corresponding pixel at multiple time points; generate a time series curve for each band: Among them, R b (t) is the reflectivity of band b at time t; this operation is performed for all bands b∈[1,N].

[0096] Use linear interpolation, spline interpolation or LOESS local weighted regression methods to calculate the sampling time point t sample The spectral reflectance at is estimated:

[0097] Reconstruct the entire remote sensing feature vector: N is the total number of bands; replace the original remote sensing features Used for sample matching.

[0098] Perform difference analysis between original remote sensing features and reconstructed remote sensing features: δ is the consistency deviation; ∈ means to prevent division by zero and is set to a minimum value; if δ is greater than the set threshold (such as 0.15), it means that the remote sensing features are seriously inconsistent with the labels.

[0099] According to the degree of deviation, the confidence of the sample label is adjusted: confidence weight setting: w = 1-δ; used for sample weighting or loss function adjustment in model training; fuzzy label generation (if it is a classification task): convert the label from hard classification (such as black and odor = 1, non-black and odor = 0) to soft label (such as [0.7, 0.3]); confidence is linked to w to alleviate the misleading effect of labels. Use reconstructed remote sensing features As model input; use the corrected labels or confidences to participate in training.

[0100] The specific steps of introducing the credibility factor for deep learning model training are as follows: using the aforementioned process to construct a hyperspectral remote sensing sample set; each sample contains: remote sensing spectral feature vector; black and odorous water quality label (classification or regression); WVSI value and SPC value.

[0101] WVSI and SPC are combined into a unified credibility score (such as sample credibility level SCL); low-credibility samples are no longer eliminated, but the score is used for weight control in training.

[0102] During the training process, the loss term of each sample is weighted according to its credibility factor; high-credibility samples have a larger weight in the loss function, and low-credibility samples have a smaller weight; this enables the model to focus on learning reliable samples and reduce noise interference.

[0103] A model architecture suitable for hyperspectral data (such as 1D-CNN, 3D-CNN, ResNet, Transformer) is selected; the input is the remote sensing spectral vector, and the label is the water quality status; the weighted loss function controls the training direction and iteratively optimizes the parameters.

[0104] The trained model is applied to the entire urban remote sensing image; the spectral features of each valid water pixel are extracted and input into the model.

[0105] The model outputs the black and odorous water quality prediction results (classification or continuous value) corresponding to each pixel; it can also output probability, confidence level, black and odorous grade, etc.

[0106] Reorganize pixel determination results into images or vector layers; visualize the distribution of black and odorous pollution in urban water bodies; and further conduct pollution level statistics, hotspot identification and decision support.

[0107] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0108] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0109] It should be understood that the term "and / or" herein is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent three situations: A exists alone, A and B exist at the same time, and B exists alone, where A and B may be singular or plural. In addition, the character " / " herein generally indicates that the objects associated with each other are in an "or" relationship, but it may also indicate an "and / or" relationship, which can be understood by referring to the context. A person of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0110] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A method for determining the water quality of urban black and odorous water bodies based on a deep learning model of hyperspectral remote sensing data, characterized by: include: Obtain water quality monitoring data and remote sensing images from several sampling points; Based on water quality monitoring data, the frequency and magnitude of water quality changes in the area where the sampling points are located are analyzed, and the sensitivity index of water body dynamic changes is calculated; Use the endmember extraction algorithm to obtain the spectral mixing degree of the remote sensing pixel corresponding to the sampling point, determine its interference degree, and calculate the confidence level of the spectral purity of the remote sensing pixel; The water body dynamic change sensitivity index and remote sensing pixel spectral purity confidence level corresponding to each sampling point are comprehensively scored to generate the sample credibility level, and low-credibility samples are eliminated; For samples in areas where the sensitivity index of water body dynamic changes is greater than the set threshold, time series images are introduced for interpolation to reconstruct the remote sensing characteristics of the sampling time points and perform confidence correction on the labels; The water body dynamic change sensitivity index and remote sensing pixel spectral purity confidence of the sample are used as weighting factors of the deep learning model loss function to train the deep learning model; Use the trained deep learning model to determine the black and odorous water quality of urban water bodies.

2. The method for determining the water quality of urban black and odorous water bodies based on a deep learning model of hyperspectral remote sensing data according to claim 1 is characterized in that: Obtaining water quality monitoring data includes: recording the sampling time during the field sampling process, using differential GPS to record the geographic coordinates of the sampling points, with a positioning error of less than 1 meter; and simultaneously collecting hydrological environmental information, including water flow velocity, distribution of surrounding sewage outlets, rainfall in the 24 hours before sampling, water color and turbidity.

3. The method for determining the water quality of urban black and odorous water bodies based on a deep learning model of hyperspectral remote sensing data according to claim 1 is characterized in that: The method for obtaining the water body dynamic change sensitivity index (WVSI) is to obtain multiple water quality measurement values ​​of the sampling point or its surrounding adjacent pixels within a continuous time period; including COD, NH3-N and DO; and calculate the change frequency of the water quality monitoring data: F is the frequency of change, N var is the number of samples with significant changes, N total is the total sample number; Calculate the change range of water quality monitoring data: A is the average change amplitude, Q i is the water quality index value of the i-th sampling in the time series, and N is the number of sequence samples; the change frequency and change amplitude of the water quality monitoring data are calculated for different indicators respectively, and the water body dynamic change sensitivity index WVSI is obtained by normalizing them and performing weighted summation.

4. The method for determining the water quality of urban black and odorous water bodies based on a deep learning model of hyperspectral remote sensing data according to claim 3 is characterized in that: The endmember extraction algorithm is used to obtain the spectral mixing degree of the remote sensing pixel corresponding to the sampling point, determine its interference degree, and calculate the confidence level of the spectral purity of the remote sensing pixel. Specifically: Use the water extraction algorithm to extract the water area in the remote sensing image; locate each sampling point to the corresponding remote sensing pixel and its surrounding pixel set for analysis; the goal is to identify the purest spectral pixel as the end member; use the spectral angle or linear mixture model to calculate the similarity between the pixel corresponding to the current sampling point and the purest water end member, the expression is: Where, is the reflectivity vector of the sampling point pixel, is the water end-member spectral vector, θ is the spectral angle in radians; Calculate the spectral purity confidence SPC of remote sensing pixels and obtain it by reverse normalization according to the spectral angle: SPC∈[0,1],θ max It is the maximum spectral angle of the water pixel in the dataset and is used for normalization.

5. The method for determining the water quality of urban black and odorous water bodies based on a deep learning model of hyperspectral remote sensing data according to claim 4 is characterized in that: After normalizing WVSI and SPC, the sample confidence score is calculated using the expression: SCL = λ·WVSI′ + (1-λ)·SPC′; where SPC′ is the confidence level of the spectral purity of the remote sensing pixel after normalization. WVSI′ is the water body dynamic change sensitivity index after normalization; SCL is the sample credibility level, ranging from [0,1]; λ is the weight coefficient.

6. The method for determining the water quality of urban black and odorous water bodies based on a deep learning model of hyperspectral remote sensing data according to claim 5 is characterized in that: According to the comprehensive score SCL, the samples are divided into different levels: High confidence level: SCL ≥ 0.8, the sample is stable and the observation is pure, and can be used directly for training; Medium credibility level: 0.6≤SCL<0.8, the sample is credible but has slight interference and can be used with a lower priority; Low confidence level: SCL < 0.6, the samples are potentially mismatched or mixed and should be eliminated or used for uncertainty modeling.

7. The method for determining the water quality of urban black and odorous water bodies based on a deep learning model of hyperspectral remote sensing data according to claim 6 is characterized in that: For samples in areas where the sensitivity index of water body dynamic changes is greater than the set threshold, time series images are introduced for interpolation to reconstruct the remote sensing characteristics of the sampling time points and perform confidence correction on the labels: Traverse all sampling points and identify their corresponding WVSI values; The samples whose WVSI exceeds the set threshold are classified into the sensitive area set; The reliability of the synchronization between the remote sensing features and the measured labels of the samples in the collection is low and needs to be processed; Collect remote sensing images of the same area within the time range before and after the sampling time point; For each high WVSI sample, extract the spectral vector of its corresponding pixel at multiple time points; generate a time series curve for each band: Among them, R b (t) is the reflectivity of band b at time t; this operation is performed for all bands b∈[1,N]; T-1, T0, T+1 are three different time points; For the sampling time point t sample The spectral reflectance at is estimated: Reconstruct the entire remote sensing feature vector: N is the total number of bands; replace the original remote sensing features Used for sample matching; Perform difference analysis between original remote sensing features and reconstructed remote sensing features: δ is the consistency deviation; ∈ indicates preventing division by zero and is set to the minimum value; According to the degree of deviation, the confidence of the sample label is adjusted, and the reconstructed remote sensing features are used. As model input, and use the corrected labels or confidences to participate in training.

Images