Water quality inversion method and device, computer device and storage medium

Abstract

The application provides a water quality inversion method and device, computer equipment and a storage medium. The method comprises: acquiring wide-area hyperspectral image data, regional multispectral image data and measured water quality parameters; correcting and spatiotemporally registering the wide-area hyperspectral image data and the regional multispectral image data, and performing feature fusion to generate a fused feature vector; based on the fused feature vector, a first type of water quality parameter is obtained through a first inversion model; based on the first type of water quality parameter and the measured water quality parameter, a second type of water quality parameter is obtained through a second inversion model; based on the first type of water quality parameter and the second type of water quality parameter, a water quality parameter spatial distribution map is obtained; and according to the water quality parameter spatial distribution map, when the measured water quality parameter is greater than a preset threshold, a warning signal is generated. Based on the application scheme, the accuracy, timeliness and automation level of water quality parameter inversion in a complex water environment can be improved.

Description

Technical Field

[0001] This application relates to the field of water environment monitoring, specifically to a water quality inversion method, apparatus, computer equipment, and storage medium. Background Technology

[0002] Remote sensing technologies (such as hyperspectral and multispectral imaging) have been applied to water quality monitoring. Water quality remote sensing monitoring mostly relies on a single platform (such as satellites, fixed observation towers, or drones), which has significant limitations: satellites have long revisit cycles and are easily affected by cloud interference, making high-frequency monitoring difficult; fixed observation towers can conduct wide-area continuous observations, but their spatial resolution and spectral information are limited, restricting accurate multi-parameter inversion; although drones can provide high-resolution data, their coverage and endurance are insufficient, failing to meet the high-frequency monitoring needs of large lakes and reservoirs.

[0003] A single remote sensing platform cannot simultaneously cover large-scale, high-resolution, and high-frequency monitoring targets, resulting in insufficient ability to capture dynamic water quality changes in complex water bodies such as lakes and reservoirs, making it difficult to achieve real-time early warning and efficient management of water pollution incidents. Summary of the Invention

[0004] This application provides a water quality inversion method, apparatus, computer equipment, and storage medium. A first aspect of this application provides a water quality inversion method, which includes:

[0005] Acquire wide-area hyperspectral image data collected by fixed observation platforms, regional multispectral image data collected by mobile observation platforms, and measured water quality parameters collected by ground monitoring points;

[0006] The wide-area hyperspectral image data and regional multispectral image data are corrected and spatiotemporally registered. The registered wide-area hyperspectral image data and the registered regional multispectral image data are then fused to generate a fused feature vector.

[0007] Based on the fused feature vectors, the first type of water quality parameters are obtained by inversion using the first inversion model;

[0008] Based on the first type of water quality parameters and the measured water quality parameters, the second type of water quality parameters are obtained by inversion through the second inversion model;

[0009] The spatial distribution map of water quality parameters is obtained by inversion based on the first type of water quality parameters and the second type of water quality parameters;

[0010] Based on the spatial distribution map of water quality parameters, an early warning signal is generated when the measured water quality parameters exceed the preset threshold.

[0011] In an optional embodiment of this application, wide-area hyperspectral image data and regional multispectral image data are corrected and spatiotemporally registered, and the registered wide-area hyperspectral image data and the registered regional multispectral image data are feature-fused to generate a fused feature vector, including:

[0012] Radiometric calibration, atmospheric correction, geometric correction, image stitching, and normalized differential water index extraction were performed on wide-area hyperspectral image data to obtain hyperspectral water reflectance data.

[0013] Radiometric calibration, atmospheric correction, geometric correction, image stitching, and normalized differential water index extraction were performed on the regional multispectral image data to obtain multispectral reflectance data.

[0014] Based on the scale-invariant feature transformation algorithm, spatiotemporal registration of hyperspectral water-leaving reflectance data and multispectral reflectance data are performed, and feature fusion is carried out to generate a fused feature vector.

[0015] In an optional embodiment of this application, based on the scale-invariant feature transform algorithm, spatiotemporal registration is performed on hyperspectral water-leaving reflectance data and multispectral reflectance data, and feature fusion is performed to generate a fused feature vector, including:

[0016] Based on the scale-invariant feature transform algorithm, spatiotemporal registration of hyperspectral water-free reflectance data and multispectral reflectance data is performed to obtain registered hyperspectral water-free reflectance data and registered multispectral reflectance data.

[0017] Based on the registered hyperspectral water-free reflectance data and the registered multispectral reflectance data, obtain the reflectance values ​​of the registered hyperspectral water-free reflectance data in the preset red band and the preset near-infrared band, as well as the reflectance values ​​of the registered multispectral reflectance data in the preset red band and the preset near-infrared band.

[0018] Based on the registered hyperspectral water-leaving reflectance data in the preset red band and the preset near-infrared band, and the registered multispectral reflectance data in the preset red band and the preset near-infrared band, the spectral difference compensation factor is calculated, as shown in the following formula:

[0019]

[0020] Where D is the spectral difference compensation factor, R HSI (λ red R represents the reflectance value of the registered hyperspectral water-free reflectance data in the preset red band. HSI (λ NIRR represents the reflectance value of the registered hyperspectral water-leaving reflectance data in the preset near-infrared band. MSI (λ red R represents the reflectance value of the registered multispectral reflectance data in the preset red band. MSI (λ NIR () represents the reflectance value of the registered multispectral reflectance data in the preset near-infrared band;

[0021] Based on the spectral difference compensation factor, the registered hyperspectral water-leaving reflectance data, and the registered multispectral reflectance data, the fused feature vector is obtained, as shown in the following formula:

[0022] F = [HSI, MSI, D]

[0023] Where F is the fused feature vector, HSI is the registered hyperspectral water-free reflectance data, MSI is the registered multispectral reflectance data, and D is the spectral difference compensation factor.

[0024] In an optional embodiment of this application, the first type of water quality parameters includes chlorophyll a concentration, total suspended solids, turbidity, and transparency; based on the fused feature vector, the first type of water quality parameters are obtained by inversion through a first inversion model, including:

[0025] The chlorophyll a concentration was obtained by inverting the fused feature vector using a random forest regression model.

[0026] The reflectance value of the preset band is calculated based on the fused feature vector, and the total suspended matter, turbidity and transparency are obtained by inversion based on the reflectance value of the preset band.

[0027] In an optional embodiment of this application, the reflectance value of a preset band is calculated based on the fused feature vector, and inversion is performed based on the reflectance value of the preset band to obtain total suspended matter, turbidity, and transparency, including:

[0028] Based on the measured water quality parameters, the inversion coefficients were obtained by calibration using the least squares method.

[0029] Inversion models for total suspended solids, turbidity, and transparency are established based on the inversion coefficients, as shown in the following formulas:

[0030]

[0031] Where Y represents total suspended solids, turbidity, or transparency, and k0 and k i Let λ be the inversion coefficients obtained by calibration using the least squares method. a , λ b Preset band;

[0032] The reflectance value of a preset band is calculated by fusing feature vectors;

[0033] Inversion is performed based on the reflectance value of the preset band, the inversion model of total suspended matter, the inversion model of turbidity, and the inversion model of transparency to obtain total suspended matter, turbidity, and transparency.

[0034] In an optional embodiment of this application, the second type of water quality parameters are obtained by inversion using a second inversion model based on the first type of water quality parameters and the measured water quality parameters, and the method further includes:

[0035] Obtain the water temperature and pH value from the measured water quality parameters;

[0036] Based on water temperature, pH value, chlorophyll a concentration, total suspended matter, turbidity, and transparency, a subset of input features is obtained by dynamic feature selection using recursive feature elimination and random forest regression models.

[0037] Based on the input feature subset, the second type of water quality parameters are obtained by inversion through the second inversion model.

[0038] In an optional embodiment of this application, a second type of water quality parameters are obtained by inverting a second inversion model based on an input feature subset, including:

[0039] Base learners are constructed based on K-nearest neighbor regression, support vector regression, and extreme gradient boosting regression models, respectively.

[0040] A meta-learner is constructed based on a pre-defined linear regression model. The prediction function of the meta-learner is shown in the following formula:

[0041]

[0042] Where Y is the second type of water quality parameter predicted by the meta-learner, k is the index value of the base learner, and w k Assign weights to the base learners to the meta-learner, f k Let X represent the prediction function of the base learner, where X is a subset of the input features;

[0043] A stacked ensemble model is constructed based on the base learner and the meta-learner to obtain the second inversion model;

[0044] Based on the input feature subset, the second type of water quality parameters are obtained by inversion through the second inversion model. The second type of water quality parameters include dissolved oxygen, permanganate index, chemical oxygen demand, ammonia nitrogen, total nitrogen, and total phosphorus.

[0045] A second aspect of this application provides a water quality inversion device, the device comprising:

[0046] The acquisition unit is used to acquire wide-area hyperspectral image data collected by fixed observation platforms, regional multispectral image data collected by mobile observation platforms, and measured water quality parameters collected by ground monitoring points.

[0047] The data processing unit is used to perform correction and spatiotemporal registration processing on wide-area hyperspectral image data and regional multispectral image data, and to perform feature fusion on the registered wide-area hyperspectral image data and the registered regional multispectral image data to generate a fused feature vector.

[0048] The first inversion unit is used to obtain the first type of water quality parameters by inverting the first inversion model based on the fused feature vector.

[0049] The second inversion unit is used to invert the second type of water quality parameters based on the first type of water quality parameters and the measured water quality parameters through the second inversion model.

[0050] The spatial distribution map generation unit is used to invert the spatial distribution map of water quality parameters based on the first type of water quality parameters and the second type of water quality parameters.

[0051] The monitoring unit is used to generate an early warning signal when the measured water quality parameters exceed a preset threshold, based on the spatial distribution map of water quality parameters.

[0052] A third aspect of this application provides a computer device, including: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above methods.

[0053] A fourth aspect of the present application provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the method as described in any of the above.

[0054] This application's embodiments effectively overcome the inherent limitations of a single remote sensing platform in water quality monitoring by comprehensively utilizing multi-source data acquisition and fusion processing from fixed observation platforms, mobile observation platforms, and ground monitoring points. First, by acquiring wide-area hyperspectral and regional multispectral images and performing correction and spatiotemporal registration, both wide-area coverage is preserved while supplementing key spatial details and spectral information, thus achieving a unification of large-scale and high-resolution data at the data level. Furthermore, feature fusion generates fused feature vectors, enhancing the representational ability of multi-scale and multispectral features and laying the foundation for subsequent accurate inversion. Subsequently, based on the fused features, the first type of water quality parameters are obtained through a first inversion model, initially establishing the mapping relationship between remote sensing features and water quality parameters. Then, combined with ground-measured water quality parameters, the inversion process is further optimized through a second inversion model to obtain more accurate second type of water quality parameters. This achieves hierarchical fusion and complementarity of multi-source information at the model level, improving the robustness and accuracy of multi-parameter inversion in complex aquatic environments. Finally, a spatial distribution map of water quality parameters is generated based on multiple inversion results, and an early warning is issued when the measured value exceeds the threshold. This realizes a closed-loop management system from wide-area monitoring to local precise identification and then to real-time early warning, which significantly enhances the ability to capture dynamic water quality changes in lakes and reservoirs and the ability to respond to pollution events in real time. Attached Figure Description

[0055] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0056] Figure 1 A schematic diagram of the system architecture of a water quality inversion method provided in one embodiment of this application;

[0057] Figure 2 A flowchart of a water quality inversion method provided in one embodiment of this application;

[0058] Figure 3 This is a schematic diagram of a hierarchical inversion model provided in one embodiment of this application;

[0059] Figure 4 A flowchart of a water quality inversion method provided in one embodiment of this application;

[0060] Figure 5 A flowchart of a water quality inversion method provided in one embodiment of this application;

[0061] Figure 6 A flowchart of a water quality inversion method provided in one embodiment of this application;

[0062] Figure 7 A flowchart of a water quality inversion method provided in one embodiment of this application;

[0063] Figure 8 A flowchart of a water quality inversion method provided in one embodiment of this application;

[0064] Figure 9 A flowchart of a water quality inversion method provided in one embodiment of this application;

[0065] Figure 10 This is a schematic diagram of the structure of a water quality inversion device provided in one embodiment of this application;

[0066] Figure 11 This is a schematic diagram of a stacked ensemble learning framework provided in one embodiment of this application;

[0067] Figure 12 This is a schematic diagram of the structure of a water quality inversion device provided in one embodiment of this application;

[0068] Figure 13 This is a schematic diagram of a computer device structure provided in one embodiment of this application. Detailed Implementation

[0069] The solutions in this application embodiment can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0070] To make the technical solutions and advantages of the embodiments of this application clearer, the exemplary embodiments of this application will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0071] The following is a brief description of the application environment of the water quality inversion method provided in the embodiments of this application:

[0072] Remote sensing technologies (such as hyperspectral and multispectral imaging) have been applied to water quality monitoring, but existing methods mostly rely on a single platform (such as satellites, fixed observation towers, or drones), which has significant limitations: satellites have long revisit cycles and are easily affected by cloud cover, making high-frequency monitoring difficult; fixed observation towers, while capable of wide-area continuous observation, have limited spatial resolution and spectral information, restricting accurate multi-parameter inversion; drones, while providing high-resolution data, have insufficient coverage and endurance, failing to meet the high-frequency monitoring needs of large lakes and reservoirs. A single remote sensing platform cannot simultaneously handle large-scale, high-resolution, and high-frequency monitoring targets, resulting in insufficient effectiveness for dynamic water quality monitoring in complex aquatic environments (such as eutrophic lakes and reservoirs).

[0073] Furthermore, for non-optically reactive water quality parameters such as dissolved oxygen (DO), total nitrogen (TN), and total phosphorus (TP), current technologies still heavily rely on manual on-site sampling and laboratory testing. This approach suffers from inherent drawbacks such as long processing times (from several hours to several days), high costs, and sparse sampling points, severely limiting the real-time response capability to pollution events. Simultaneously, traditional inversion algorithms lack robustness in variable weather and complex optical environments, restricting the potential of remote sensing technology to reduce reliance on manual sampling and achieve rapid estimation of water quality parameters across the entire spectrum.

[0074] This application proposes a solution: by comprehensively utilizing multi-source data from fixed observation platforms, mobile observation platforms, and ground monitoring points for collaborative acquisition and fusion processing, the inherent limitations of a single remote sensing platform in water quality monitoring are effectively overcome. First, by acquiring wide-area hyperspectral and regional multispectral images and performing correction and spatiotemporal registration, both wide-area coverage is preserved while supplementing key spatial details and spectral information, thus achieving a unification of large-scale and high-resolution data at the data level. Furthermore, by generating fused feature vectors through feature fusion, the representation capability of multi-scale and multispectral features is enhanced, laying the foundation for subsequent accurate inversion. Subsequently, based on the fused features, the first type of water quality parameters are obtained through the first inversion model, initially establishing the mapping relationship between remote sensing features and water quality parameters. Then, combined with ground-measured water quality parameters, the inversion process is further optimized through the second inversion model to obtain more accurate second type of water quality parameters. This achieves hierarchical fusion and complementarity of multi-source information at the model level, improving the robustness and accuracy of multi-parameter inversion in complex aquatic environments. Finally, a spatial distribution map of water quality parameters is generated based on multiple inversion results, and an early warning is issued when the measured value exceeds the threshold. This realizes a closed-loop management system from wide-area monitoring to local precise identification and then to real-time early warning, which significantly enhances the ability to capture dynamic water quality changes in lakes and reservoirs and the ability to respond to pollution events in real time.

[0075] Combination Figure 1 - Figure 9 The water quality inversion method provided in the embodiments of this application will be described in detail.

[0076] Please see Figure 4 The following embodiments use the aforementioned water quality inversion system as the execution subject, applying the method provided in the embodiments of this application to the aforementioned water quality inversion system. Figure 4 As shown, the water quality inversion method provided in this application includes the following steps S10-S60:

[0077] S10: Acquire wide-area hyperspectral image data collected by fixed observation platforms, regional multispectral image data collected by mobile observation platforms, and measured water quality parameters collected by ground monitoring points.

[0078] It should be noted that, referring to Figure 1Fixed observation platforms refer to remote sensing platforms permanently installed in specific locations (such as monitoring towers or building rooftops). Their function is to provide large-scale, continuous coverage observation data, with a wide-area continuous monitoring radius that can be set to greater than or equal to 5 km. Wide-area hyperspectral image data is acquired by hyperspectral cameras on fixed platforms, providing continuous narrow-band, wide-coverage image data to support refined water quality parameter identification. Mobile observation platforms refer to flexibly movable platforms (such as drones or unmanned vessels), providing high spatial resolution mobile observation data for specific areas, with a resolution that can be set to less than or equal to 10 cm. Regional multispectral image data is acquired by multispectral sensors on mobile platforms, providing several discrete wide-band, high-spatial-resolution image data to supplement high spatial resolution detail information. Ground monitoring points are on-site sampling points deployed in water bodies. The measured water quality parameters collected, such as chlorophyll a, turbidity, dissolved oxygen, total nitrogen, and total phosphorus, serve as baseline ground truth data for subsequent model training, validation, and final early warning, ensuring the accuracy of the inversion results.

[0079] Understandably, the coordinated deployment of three heterogeneous monitoring platforms—fixed, mobile, and ground-based—aims to acquire complementary multi-source data from the source. Fixed platforms ensure the wide coverage and continuity of monitoring, mobile platforms compensate for the insufficient spatial resolution of fixed platforms, and ground-based data provides an indispensable field verification and calibration basis for remote sensing inversion. Together, they lay the data foundation for addressing the inherent limitations of a single platform in terms of range, resolution, and frequency.

[0080] Specifically, fixed observation platforms can be monitoring towers equipped with hyperspectral imaging systems, while mobile observation platforms can be drones equipped with multispectral cameras. GPS timing technology ensures high synchronization of data acquisition times for both types of platforms (e.g., deviations controlled within 30 minutes) to guarantee spatiotemporal data consistency. Ground monitoring points will conduct simultaneous sampling and laboratory analysis in typical areas of the water body according to a pre-defined plan.

[0081] S20, the wide-area hyperspectral image data and the regional multispectral image data are corrected and spatiotemporally registered, and the registered wide-area hyperspectral image data and the registered regional multispectral image data are feature fused to generate a fused feature vector.

[0082] It should be noted that, referring to Figure 2 Correction can include atmospheric correction and orthorectification. Atmospheric correction can eliminate the influence of atmospheric scattering and absorption on the signal, ultimately obtaining water reflectance data that truly reflects the characteristics of the water body. Orthorectification can perform geometric processing on remote sensing images, eliminating image distortion and displacement caused by factors such as terrain undulation and sensor tilt, thereby generating an image with a uniform scale and consistent image quality. Figure 1 Such accurate orthophotos.

[0083] Spatiotemporal registration refers to the process of unifying image data (wide-field hyperspectral and regional multispectral) from different platforms, times, perspectives, and resolutions under the same geometric space and temporal reference. This is typically achieved using feature point matching algorithms (such as SIFT), which ensure a one-to-one correspondence between geographic locations of pixels from different source data, providing a foundation for subsequent fusion. The two types of registered data are then combined at the pixel level. The fusion feature vector is the representation of the fused data. For example, the spectral channels of hyperspectral data, the bands of multispectral data, and compensation factors characterizing the differences in spectral response between the two can be combined into a multidimensional feature vector. Its purpose is to comprehensively utilize wide-field spectral information and high spatial resolution details to form richer and more robust feature inputs.

[0084] Understandably, the correction process ensured the physical consistency and comparability of the data; the high-precision spatiotemporal registration solved the spatial registration problem between heterogeneous data, creating conditions for pixel-level fusion; and finally, through feature fusion, the spectral advantages of hyperspectral and spatial advantages of multispectral were organically combined to generate a fusion feature vector with more comprehensive information dimensions and stronger anti-interference ability, thereby significantly improving the feature expression ability and generalization performance of the subsequent inversion model.

[0085] Specifically, the calibration process can be completed using the radiometric calibration and atmospheric correction modules (such as FLAASH and QUAC) in professional remote sensing software such as ENVI and ERDAS. Spatiotemporal registration can be achieved through geometric correction using SIFT feature point matching combined with affine or polynomial transformation models.

[0086] S30, Based on the fused feature vector, the first type of water quality parameters are obtained by inversion through the first inversion model.

[0087] It should be noted that the first inversion model refers to a model specifically designed to directly estimate parameters closely related to the optical properties of water bodies from remote sensing reflectance characteristics. The first type of water quality parameters are optically active parameters, that is, parameters whose concentration changes directly cause detectable changes in the surface optical properties of water bodies (such as reflectance spectra), including chlorophyll a concentration (Chla), total suspended solids (TSM), turbidity (TURB), and transparency (SDD).

[0088] Understandably, since the first type of water quality parameters have a direct physical correlation or strong statistical correlation with the spectral signal, they can be effectively inverted directly from the feature vector that integrates wide-area spectral information and high spatial details. This achieves the first transformation from remote sensing data to basic water quality parameters, providing an inversion basis for subsequent indirect inversion of non-optical active parameters (the second type of water quality parameters) that are not directly related to the optical signal.

[0089] Specifically, the first inversion model can take several forms. For chlorophyll a concentration, machine learning models such as random forests and support vector regression can be used to handle high-dimensional nonlinear fusion features; for total suspended matter and turbidity, semi-empirical models based on specific band ratio combinations can be used; for transparency, an empirical relationship model between transparency and total suspended matter and turbidity can be established. Model parameters need to be trained or calibrated using ground-based measured data.

[0090] S40, based on the first type of water quality parameters and the measured water quality parameters, the second type of water quality parameters are obtained by inversion through the second inversion model.

[0091] It should be noted that, referring to Figure 3 The second inversion model is used to estimate parameters that have no direct or weak relationship with the surface optical properties of water bodies. The second category of water quality parameters refers to non-optically active parameters, such as dissolved oxygen (DO), chemical oxygen demand (COD), and permanganate index (COD). Mn These parameters include ammonia nitrogen (NH3-N), total nitrogen (TN), and total phosphorus (TP). These parameters cannot be directly and stably retrieved from spectral reflectance, but their concentrations in water bodies often have intrinsic biochemical and physical relationships with optically active parameters (Class I water quality parameters) and environmental factors (such as water temperature and pH, derived from measured data).

[0092] Understandably, using the first type of water quality parameters and environmental factors (such as water temperature and pH value) monitored synchronously on the ground as input features avoids the dilemma of forcibly establishing an unreliable model between weak signals and non-optically active parameters, and greatly expands the parameter range of remote sensing water quality monitoring.

[0093] Specifically, the second inversion model employs machine learning or ensemble learning models (such as extreme gradient boosting regression models and stacked ensemble models) to capture the complex nonlinear relationship between input features and target parameters. Before inputting the model, feature selection methods (such as recursive feature elimination, RFE) can be used to screen the subset most important for specific second-class water quality parameters from candidate features such as chlorophyll a concentration, total suspended solids, turbidity, transparency, water temperature, and pH value, thereby improving model efficiency and generalization ability.

[0094] S50, a spatial distribution map of water quality parameters is obtained by inversion based on the first type of water quality parameters and the second type of water quality parameters.

[0095] It should be noted that the spatial distribution map of water quality parameters is a geographic information visualization product. It renders the values ​​of the first and second categories of water quality parameters obtained from the inversion at each pixel location on a map using different colors or symbols according to their geographic coordinates, thereby intuitively and continuously displaying the spatial distribution and changing trends of each parameter in the entire water area.

[0096] Understandably, integrating single, point-based inversion values ​​into a map that can macroscopically reflect the spatial heterogeneity of water quality transforms the discrete point results obtained from the preceding inversion into continuous spatial information that can support decision-making. This allows managers to clearly identify polluted areas, pollution levels, and spatial correlations between different parameters, thereby greatly improving the intuitiveness and usability of monitoring results.

[0097] Specifically, spatial distribution maps of water quality parameters can be generated in GIS (Geographic Information System) software or programming libraries (such as ArcGIS, QGIS, or Python's GDAL / rasterio library). By constructing a spatial raster layer from the inversion results (the value of each pixel and its latitude and longitude coordinates) and assigning color gradients to values ​​of different concentrations, a distribution map in the form of a heatmap can be generated.

[0098] For example, the chlorophyll a concentration values ​​obtained from the inversion of all pixels are used to generate a chlorophyll a spatial distribution heat map covering the entire reservoir. In the map, the red area represents the high concentration area (where algal blooms may occur), and the blue area represents the low concentration area, thus clearly showing the distribution range of algal blooms.

[0099] S60, based on the spatial distribution map of the water quality parameters, when the measured water quality parameters are greater than a preset threshold, an early warning signal is generated.

[0100] It should be noted that the preset threshold is a critical concentration value set for a specific water quality parameter. When this value is exceeded, the water body is considered to be in a potentially unhealthy state or at risk of pollution. This threshold can be dynamically adjusted based on the water body's function, seasonal changes, historical data, and water environment standards. The early warning signal system automatically generates an alarm message to alert management personnel to potential water quality anomalies or pollution events. Based on the generated spatial distribution map of water quality parameters, the system automatically and quickly scans the entire water body. Once any measured key parameter (or parameter verified by inversion) in any area is found to exceed the safety threshold, an early warning is immediately triggered. This significantly shortens the time from anomaly detection to response decision-making, enabling real-time or near real-time monitoring and efficient management of water pollution events, and solving the problem of delayed response in traditional methods.

[0101] Specifically, the early warning logic can be implemented programmatically, iterating through all pixel values ​​on the spatial distribution map of water quality parameters, or focusing on the statistical values ​​(such as maximum and average values) of specific sensitive areas, and comparing them with preset thresholds in the database. If the threshold is exceeded, early warning signals such as emails, SMS messages, and platform pop-ups are automatically triggered via API interface.

[0102] For example, refer to Figure 2The system monitors the chlorophyll a concentration in the spatial distribution map of water quality parameters. If it finds that the average chlorophyll a concentration in a certain lake bay area exceeds the eutrophication threshold set for summer (e.g., 20 μg / L) for several consecutive time periods, the system will automatically generate an early warning message and send it to the reservoir management personnel via SMS, prompting "Algal bloom may occur in a certain lake bay area, please pay attention".

[0103] In this embodiment, the inherent limitations of a single remote sensing platform in water quality monitoring are effectively overcome by comprehensively utilizing multi-source data acquisition and fusion processing from fixed observation platforms, mobile observation platforms, and ground monitoring points. First, by acquiring wide-area hyperspectral and regional multispectral images and performing correction and spatiotemporal registration, both wide-area coverage is preserved while supplementing key spatial details and spectral information, thus achieving a balance between large-scale and high-resolution data at the data level. Furthermore, feature fusion generates fused feature vectors, enhancing the representational ability of multi-scale and multispectral features and laying the foundation for subsequent accurate inversion. Subsequently, based on the fused features, the first type of water quality parameters are obtained through the first inversion model, initially establishing the mapping relationship between remote sensing features and water quality parameters. Then, combined with ground-measured water quality parameters, the inversion process is further optimized through the second inversion model to obtain more accurate second type of water quality parameters. This achieves hierarchical fusion and complementarity of multi-source information at the model level, improving the robustness and accuracy of multi-parameter inversion in complex aquatic environments. Finally, a spatial distribution map of water quality parameters is generated based on multiple inversion results, and an early warning is issued when the measured value exceeds the threshold. This realizes a closed-loop management system from wide-area monitoring to local precise identification and then to real-time early warning, which significantly enhances the ability to capture dynamic water quality changes in lakes and reservoirs and the ability to respond to pollution events in real time.

[0104] Please see Figure 5 This is a flowchart of a water quality inversion method provided in one embodiment of this application. Figure 5 As shown, the method in this embodiment may include the following steps S201-S203, and steps S201-S203 may be used as a method for... Figure 4 A refinement of step S20 in the illustrated embodiment.

[0105] S201, Perform radiometric calibration, atmospheric correction, geometric correction, image stitching, and normalized differential water index extraction on the wide-area hyperspectral image data to obtain hyperspectral water reflectance data.

[0106] S202, Radiometric calibration, atmospheric correction, geometric correction, image stitching, and normalized differential water index extraction are performed on the multispectral image data of the region to obtain multispectral reflectance data;

[0107] S203, based on the scale-invariant feature transformation algorithm, performs spatiotemporal registration on the hyperspectral water-leaving reflectance data and the multispectral reflectance data, and performs feature fusion to generate a fused feature vector.

[0108] It should be noted that radiometric calibration is the process of converting the raw digital quantization values ​​(DN values) recorded by the sensor into physically meaningful apparent radiance values. Its purpose is to eliminate the influence of differences in sensor response and establish a quantitative relationship between image data and entrance pupil radiance, forming the basis for subsequent quantitative inversion. Atmospheric correction is the process of eliminating the influence of atmospheric effects such as scattering and absorption by atmospheric molecules and aerosols on the signals received by satellite or airborne sensors, thereby obtaining the true surface reflectance of ground objects or, for water bodies, the reflectance upon leaving the water. Its purpose is to ensure that the image data accurately reflects the optical characteristics of water bodies, rather than atmospheric conditions. Geometric correction eliminates image geometric distortions caused by sensor attitude, terrain undulations, and the curvature of the Earth, conforming it to the map projection system. Its purpose is to ensure the spatial geometric accuracy of the image, providing a basis for multi-source data registration and spatial analysis. Based on geometric correction, orthorectification can also be performed to eliminate displacement caused by topographic relief (DEM). Image stitching is the process of combining multiple remote sensing images with overlapping areas into a complete, seamless image covering the entire study area. Its purpose is to expand the spatial coverage of a single monitoring session. The Normalized Difference Water Index (NDWI) is an index constructed using near-infrared and green light band reflectance to enhance water body information and suppress non-water background information. The formula for calculating the NDWI is NDWI = (Green - NIR) / (Green + NIR), where Green represents the green light band (typically about 0.52-0.60 μm) and NIR represents the near-infrared band (typically about 0.76-0.90 μm). It can automatically and accurately extract water areas from images, eliminate interference from land, vegetation, etc., and ensure that subsequent processing is only applied to valid water pixels.

[0109] Specifically, this series of preprocessing operations can be completed in professional remote sensing software such as ENVI and ERDAS Imagine. Radiometric calibration requires the use of calibration coefficients provided by the sensor. Atmospheric correction can be performed using models such as FLAASH, 6S, and QUAC, depending on data conditions and accuracy requirements. Geometric correction can employ precise correction based on ground control points or directly use the sensor's built-in geometric coarse correction products and orthorectification tools. Stitching can be performed using the software's built-in seamless stitching tool. NDWI calculation and threshold segmentation can be automated.

[0110] Similarly, radiometric calibration, atmospheric correction, geometric correction, image stitching, and normalized difference water index extraction are performed on regional multispectral image data to obtain multispectral reflectance data. However, it is important to note the differences in sensors carried by mobile observation platforms; for example, the radiometric calibration coefficients of a UAV's multispectral camera may be provided by its manufacturer. Atmospheric correction may more often employ Quick Atmospheric Correction (QUAC) or methods based on simple scattering models. Geometric correction typically relies on the high-precision POS system (POS data) onboard the UAV to directly generate orthorectified images, or on fine correction using a small number of ground control points.

[0111] It should be noted that scale-invariant feature transform (SIN) is a computer vision algorithm used for image feature detection and description. Its key feature is invariance to image rotation, scaling, and brightness changes, as well as stability against viewpoint changes, affine transformations, and noise. It automatically finds and matches corresponding feature points between two images (hyperspectral and multispectral data) from different sources, with different resolutions, and captured from different angles. Spatiotemporal registration, based on the matched feature point pairs, calculates a spatial transformation model (such as affine or polynomial transformation) to transform the coordinate system of one image (usually a multispectral image) to the coordinate system of another image (usually a hyperspectral image), thereby achieving precise spatial alignment of the two images while ensuring temporal consistency (through time synchronization during acquisition). Its purpose is to solve the geometric differences such as translation, rotation, and scaling between images caused by platform heterogeneity, and it is a prerequisite for achieving pixel-level feature fusion.

[0112] Specifically, SIFT feature extraction and matching functions in software or programming libraries such as OpenCV and ENVI can be used. After matching, a transformation model (such as an affine transformation matrix) is calculated, and this model is used to resample the multispectral reflectance data (usually the image to be registered) to align it with the grid of the hyperspectral water-free reflectance data. Feature fusion can be achieved by concatenating the registered hyperspectral band data and multispectral band data along the channel dimension to form a longer feature vector.

[0113] For example, preprocessed hyperspectral water-leaving reflectance (HSI) data and multispectral multispectral reflectance (MSI) data are input into the algorithm. The SIFT algorithm first detects and describes thousands of feature points on the HSI and MSI images respectively; then, it finds matching point pairs in the two images by calculating the Euclidean distance between the feature descriptors; after removing mismatched point pairs, it calculates an optimal affine transformation matrix using the correct matching points; it uses this matrix to resample the MSI image to make it spatially perfectly aligned with the HSI image; finally, for each pixel position in the HSI image, it combines the spectral reflectance values ​​of all bands of the HSI image with the reflectance values ​​of all bands of the corresponding MSI image pixel to form a fused feature vector.

[0114] In this embodiment, radiometric calibration, atmospheric correction, geometric correction, stitching, and water body extraction are first performed to ensure the physical accuracy and geometric precision of both wide-area hyperspectral and regional multispectral image data, focusing on water targets. Then, the SIFT algorithm, with its scale-invariant properties, is used to achieve high-precision spatiotemporal registration of the two types of data. Finally, feature fusion organically combines the spectral advantages of hyperspectral data with the spatial advantages of multispectral data, generating a more comprehensive and robust fused feature vector. This improves the accuracy and reliability of subsequent water quality parameter inversion in terms of spatial detail resolution and adaptability to complex environments, effectively solving the limitation of a single data source in terms of information dimensions.

[0115] Please see Figure 6 Here is a flowchart of a water quality inversion method provided in one embodiment of this application, as shown below. Figure 6 As shown, the method in this embodiment may include the following steps S2031-S2034, and steps S2031-S2034 may be used as a method for... Figure 5 A refinement of step S203 in the illustrated embodiment.

[0116] S2031, Based on the scale-invariant feature transform algorithm, the hyperspectral water-free reflectance data and the multispectral reflectance data are spatiotemporally registered to obtain the registered hyperspectral water-free reflectance data and the registered multispectral reflectance data;

[0117] S2032, Based on the registered hyperspectral water-free reflectance data and the registered multispectral reflectance data, obtain the reflectance values ​​of the registered hyperspectral water-free reflectance data in a preset red band and in a preset near-infrared band, as well as the reflectance values ​​of the registered multispectral reflectance data in a preset red band and in a preset near-infrared band.

[0118] S2033, based on the reflectance values ​​of the registered hyperspectral water-leaving reflectance data in the preset red band and the preset near-infrared band, and the reflectance values ​​of the registered multispectral reflectance data in the preset red band and the preset near-infrared band, calculate the spectral difference compensation factor, as shown in the following formula:

[0119]

[0120] Where D is the spectral difference compensation factor, R HSI (λ red R represents the reflectance value of the registered hyperspectral water-leaving reflectance data in the preset red light band. HSI (λ NIR R represents the reflectance value of the registered hyperspectral water-leaving reflectance data in the preset near-infrared band. MsI (λ red R represents the reflectance value of the registered multispectral reflectance data in the preset red band. MSI (λ NIR The value of the registered multispectral reflectance data in the preset near-infrared band is denoted as .

[0121] S2034, based on the spectral difference compensation factor, the registered hyperspectral water-leaving reflectance data, and the registered multispectral reflectance data, the fused feature vector is obtained, as shown in the following formula:

[0122] F = [HSI, MSI, D]

[0123] Where F is the fused feature vector, HSI is the registered hyperspectral water-free reflectance data, MSI is the registered multispectral reflectance data, and D is the spectral difference compensation factor.

[0124] It should be noted that the Scale Invariant Feature Transform (SIFT) algorithm automatically detects stable keypoints and generates feature descriptors in two remote sensing images (i.e., hyperspectral water-leaving reflectance data and multispectral reflectance data) from different sources, with different resolutions and shooting angles, thereby achieving high-precision feature point matching between the two images. Spatiotemporal registration involves resampling and geometrically transforming the multispectral reflectance data (usually the image to be registered) to ensure that each pixel is spatially aligned precisely with its corresponding pixel in the hyperspectral water-leaving reflectance data, while ensuring that both have the same temporal reference (guaranteed through acquisition synchronization). The registered hyperspectral water-leaving reflectance data and the registered multispectral reflectance data are two sets of data that have eliminated relative geometric distortions and are in the same spatial reference frame. This can be achieved using appropriate tools from programming libraries (such as OpenCV) or remote sensing processing software (such as ENVI). First, SIFT feature point detection and descriptor generation are performed on the two images respectively. Then, algorithms such as brute-force matching or FLANN matcher are used for preliminary feature point matching. Next, algorithms such as RANSAC are used to remove incorrect matching pairs to obtain an accurate set of matching points. Finally, the transformation matrix is ​​calculated using these correct matching points, and the multispectral image is resampled to align it with the hyperspectral image.

[0125] It should be noted that the preset red light band and preset near-infrared band refer to two specific wavelength ranges selected in advance based on the sensor's band settings and the optical characteristics of the water. For most water color sensors, the red light band is typically set to 670nm, and the near-infrared band is typically set to 840nm. Preset bands that are sensitive to water composition and are comparable across different sensors are selected for subsequent calculations. The reflectance value is extracted from registered hyperspectral water reflectance data and registered multispectral reflectance data. The spectral reflectance value at a specified band and pixel location represents the optical characteristics of the water at that point in the corresponding band.

[0126] It is understandable that the red and near-infrared bands were chosen because these two bands are very sensitive to changes in the concentration of suspended sediment, chlorophyll and other components in water bodies, and are the core bands for calculating various water indices (such as NDWI). Therefore, the differences in the spectral response of these bands can better represent the overall differences between different sensor systems.

[0127] It should be noted that the spectral difference compensation factor D is a scalar factor used to address the systematic differences in reflectance values ​​between quantized and registered hyperspectral and multispectral data at the same geographical location, water body, and spectral band. The spectral difference compensation factor can be normalized to eliminate the influence of lighting conditions and the intensity of water reflection itself, thereby enhancing their comparability across different regions and times.

[0128] It should be noted that the fused feature vector F is a multi-dimensional vector formed by combining features from different sources and with different properties at the pixel level. It connects the registered hyperspectral water reflectance (HSI) data, the registered multispectral reflectance (MSI) data, and the spectral difference compensation factor (D), collectively constituting a feature vector representing the comprehensive characteristics of that pixel. By fusing the feature vector, the effective information of the original data is preserved to the maximum extent. At the same time, the introduction of the difference compensation factor enhances the model's fault tolerance and generalization performance, generating a feature representation with comprehensive information dimensions and strong anti-interference capabilities, providing data support for subsequent complex water quality parameter inversion tasks.

[0129] In this embodiment, high-precision spatiotemporal registration between heterogeneous remote sensing data is first achieved through a scale-invariant feature transform algorithm, solving the problem of unifying the spatial reference of observation data from multiple platforms. Then, by selecting the red and near-infrared bands, which are sensitive to water characteristics, precise data input is prepared for quantifying sensor differences. Next, a spectral difference compensation factor D is calculated using a designed normalized difference formula, capturing and quantifying the inherent spectral response differences between different sensor systems. Finally, by fusing hyperspectral water reflectance data, multispectral reflectance data, and the spectral difference compensation factor D into a joint feature vector, not only is the complementary advantage of wide-area spectral information, high-resolution spatial information, and system difference information achieved in the information dimension, but it also provides the subsequent inversion model with the ability to perceive and adapt to system errors in the data source at the feature level. This improves the quality and intelligence level of multi-source data fusion, enabling the final inversion model to extract water quality information more robustly and accurately from complex multi-source data, effectively overcoming the problem of decreased inversion accuracy caused by sensor heterogeneity.

[0130] Please see Figure 7 This is a flowchart illustrating a water quality inversion method provided in one embodiment of this application. The first type of water quality parameters includes chlorophyll a concentration, total suspended solids, turbidity, and transparency; such as... Figure 7 As shown, the method in this embodiment may include the following steps S301-S302, and steps S301-S302 may be used as a method for... Figure 4 A refinement of step S30 in the illustrated embodiment.

[0131] S301, the fusion feature vector is inverted using a random forest regression model to obtain the chlorophyll a concentration;

[0132] S302, calculate the reflectance value of the preset band based on the fused feature vector, and perform inversion based on the reflectance value of the preset band to obtain the total suspended matter, the turbidity and the transparency.

[0133] It should be noted that the Random Forest Regression model is an ensemble learning algorithm that improves prediction accuracy and controls overfitting by constructing multiple decision trees and averaging their predictions. The spectral response of chlorophyll a exhibits a non-linear relationship with its concentration and is susceptible to interference from other water quality parameters (such as suspended solids and yellow substances). The Random Forest Regression model can automatically extract deep, non-linear patterns and information related to chlorophyll a concentration from complex fused feature vectors without requiring manual pre-specification of specific bands or indices. This avoids the poor universality of traditional empirical models in complex water bodies and significantly improves the accuracy and robustness of chlorophyll a concentration retrieval under varying environments.

[0134] Specifically, in the random forest regression model, the optimal combination of hyperparameters, including the number of decision trees (n), is determined by grid search combined with k-fold cross-validation (k=5). estimators =500) and the maximum depth of a single tree (max depth =15). Where, n estimators =500 indicates that the ensemble contains 500 decision trees. Increasing the number of trees usually improves model stability but increases computational cost. 500 is an initial value set based on experience; in practical applications, it can be further optimized through grid search. max depth =15 is used to limit the maximum growth depth of a single tree, which helps to suppress model overfitting. This value is suitable for medium-sized datasets, and the specific optimal value needs to be determined based on the cross-validation results.

[0135] It should be noted that the reflectance value of the preset band is the reflectance value extracted from the fused feature vector at a specific wavelength or band center. The preset band can be determined in advance based on the optical properties of the water body, and is the combination of bands most sensitive to changes in total suspended matter and turbidity concentration, often involving blue, green, red, and near-infrared bands. Total suspended matter refers to the total concentration of inorganic and organic particulate matter suspended in water. An increase in its concentration usually systematically increases the reflectance of the water body in all bands, especially in the near-infrared band. Turbidity is a measure of the scattering effect of water on light, and is closely related to the concentration, particle size distribution, and composition of suspended particulate matter. Its spectral response is similar to that of total suspended matter but focuses on optical effects. Transparency is usually expressed as Sage disk depth and is a parameter that measures the light transmittance of water, mainly controlled by the concentration of suspended matter and colored dissolved organic matter.

[0136] Understandably, unlike the complex characteristics of chlorophyll a, there is a strong, approximately linear statistical relationship between the concentrations of total suspended matter and turbidity and the reflectance or its ratio in specific wavelength bands (especially the near-infrared band). Therefore, a semi-empirical log-linear model based on band ratios is adopted to improve computational efficiency while ensuring accuracy. For transparency (SDD), there is no direct linear relationship with spectral reflectance, but it is mainly controlled by the physical properties of suspended matter concentration (TSM) and turbidity (TURB). Therefore, an empirical model of transparency in relation to suspended matter concentration and turbidity is established.

[0137] Specifically, reflectance values ​​corresponding to the HSI or MSI portion are extracted from the fused feature vector F of each pixel according to a preset wavelength. Then, for total suspended matter and turbidity, the extracted reflectance values ​​are substituted into their respective preset semi-empirical model formulas (usually a linear combination of band ratios) for calculation. The model coefficients can be pre-determined using ground-based measured data through least squares fitting. For transparency, the inversion process is divided into two steps: first, TSM and TURB are inverted using the aforementioned method; then, these two inversion results are used as inputs and substituted into another preset empirical model (usually a log-linear relationship) with TSM and TURB as independent variables to calculate SDD.

[0138] In this embodiment, differentiated optimal inversion strategies were adopted based on the different optical response characteristics within the first category of water quality parameters. For chlorophyll a concentration, which exhibits complex spectral responses and significant nonlinear characteristics, a powerful random forest regression model was used to process high-dimensional fusion features, fully uncovering its implicit complex patterns and ensuring inversion accuracy and model robustness. For total suspended matter and turbidity, which have strong statistical correlations with reflectance in specific bands, a computationally efficient semi-empirical band ratio model was employed, significantly improving inversion efficiency while maintaining accuracy, meeting the needs of wide-ranging applications. For transparency, which is mainly controlled by physical processes and has no direct linear relationship with the spectrum, indirect estimation was performed using the already inverted total suspended matter and turbidity results, avoiding the dilemma of establishing unreliable direct models. Through a hierarchical and methodological inversion system, accurate and efficient simultaneous acquisition of multiple optically active parameters was achieved, providing a reliable and comprehensive input data foundation for subsequent indirect inversion of non-optically active parameters.

[0139] Please see Figure 8 Here is a flowchart of a water quality inversion method provided in one embodiment of this application, as shown below. Figure 8 As shown, the method in this embodiment may include the following steps S3021-S3024, and steps S3021-S3024 may be used as a method for... Figure 7 A refinement of step S302 in the illustrated embodiment.

[0140] S3021, Based on the measured water quality parameters, the inversion coefficients are obtained by calibration using the least squares method;

[0141] S3022, Based on the inversion coefficients, establish inversion models for the total suspended matter, the turbidity, and the transparency, as shown in the following formulas:

[0142]

[0143] Where Y represents total suspended solids, turbidity, or transparency, and k0 and k i Let λ be the inversion coefficients obtained by calibration using the least squares method. a , λ b Preset band;

[0144] S3023, Calculate the reflectance value of the preset band using the fused feature vector;

[0145] S3024, perform inversion based on the reflectivity value of the preset band, the inversion model of the total suspended matter, the inversion model of the turbidity, and the inversion model of the transparency to obtain the total suspended matter, the turbidity, and the transparency.

[0146] It should be noted that the measured water quality parameters are the concentrations or measurements of total suspended solids (TSM), turbidity (TURB), and transparency (SDD) obtained synchronously from ground monitoring points, representing the true state of the water body. These values ​​are used as true values ​​for model training and parameter calibration. The least squares method finds the best function fit for the data by minimizing the sum of squared errors. Its specific application is linear regression, aiming to find a set of inversion coefficients (k0, k...). i This minimizes the overall deviation between the natural logarithm of the water quality parameter values ​​predicted based on remote sensing reflectance characteristics (such as band ratio) and the natural logarithm of the measured values.

[0147] Understandably, while semi-empirical models are based on physical principles (such as band ratios partially eliminating the effects of illumination and water surface fluctuations), their specific mathematical relationships (coefficients) can vary depending on different sensors and aquatic environments. By using the least squares method and calibrating the model coefficients with field measurement data simultaneously collected in the current study area, the relationship between spectral and water quality parameters under the current specific environment can be accurately captured. This significantly improves the model's prediction accuracy and applicability in this region, avoiding the large biases that may result from directly using an uncalibrated general model.

[0148] It should be noted that the inversion model is a semi-empirical log-linear model:

[0149]

[0150] Where Y represents total suspended solids, turbidity, or transparency, and k0 and k i Let λ be the inversion coefficients obtained by calibration using the least squares method. a , λ b This is a preset band. ln(Y) represents taking the natural logarithm of the water quality parameter concentration Y to make the response variable more consistent with a normal distribution, thus satisfying the assumption of linear regression; The ratio is the band ratio. Using the ratio calculation can eliminate multiplicative noise caused by solar altitude angle, atmospheric conditions, water surface fluctuations, etc., to a certain extent, and enhance the true correlation between the signal and water quality parameters.

[0151] Specifically, the preset wavelengths can be determined in advance based on the optical properties of the water body, using combinations of wavelengths most sensitive to changes in total suspended solids and turbidity concentration, such as those commonly involving blue, green, red, and near-infrared bands. Preferably, taking the 490nm, 560nm, 665nm, and 705nm wavelengths as examples, the inversion model can be constructed as follows:

[0152] The inversion model for turbidity (TURB) is as follows:

[0153]

[0154] The inversion model for total suspended solids (TSM) is as follows:

[0155]

[0156] The inversion model for transparency (SDD) is as follows:

[0157] ln(SDD)=c0+c1·ln(TSM)+c2·ln(TURB)

[0158] Among them, R 705 R 655 R 560 R 490 The reflectance values ​​are for the 490nm, 560nm, 665nm, and 705nm wavelength bands, respectively, with coefficient α. i b i c i (i = 0, 1, 2) are the inversion coefficients obtained by calibration using the least squares method.

[0159] Transparency (SDD) has no direct linear relationship with spectral reflectance, but is mainly controlled by the physical properties of suspended matter concentration (TSM) and turbidity (TURB). Therefore, an empirical model of transparency in relation to suspended matter concentration and turbidity is established.

[0160] In this embodiment, the semi-empirical model is precisely calibrated using ground-based measured data via the least squares method, ensuring a high degree of adaptability between the inversion model and the current aquatic environment characteristics. Furthermore, a log-linear inversion model based on band ratios is established, leveraging the noise suppression capabilities of ratio operations and the data distribution improvement effects of logarithmic transformation to enhance the model's robustness and physical meaning. Subsequently, the preset band reflectance required by the model is accurately extracted from the fused feature vectors, providing accurate input for model execution. Finally, by substituting the feature values ​​into the calibrated model for rapid calculation, efficient and batch inversion of total suspended matter, turbidity, and transparency is achieved.

[0161] Please see Figure 9 Here is a flowchart of a water quality inversion method provided in one embodiment of this application, as shown below. Figure 9 As shown, the method in this embodiment may include the following steps S401-S403, and steps S401-S403 may be used as a method for... Figure 4 A refinement of step S40 in the illustrated embodiment.

[0162] S401, Obtain the water temperature and pH value from the measured water quality parameters;

[0163] S402, based on the water temperature, pH value, chlorophyll a concentration, total suspended matter, turbidity, and transparency, a subset of input features is obtained by dynamically filtering features using recursive feature elimination and random forest regression models;

[0164] S403, Based on the input feature subset, the second type of water quality parameters are obtained by inversion through the second inversion model.

[0165] It should be noted that water temperature (Temp) and pH values ​​were collected and measured simultaneously from ground monitoring points. These values ​​are correlated with Class II water quality parameters (non-optically active parameters), including dissolved oxygen (DO), total nitrogen (TN), and total phosphorus (TP). Recursive Feature Elimination (RFE) is a feature selection algorithm that recursively eliminates the least important features until a specified number of features or an importance threshold is reached. The base estimator (random forest regression model) is trained on all features and calculates the importance of each feature. Based on the importance of each feature, one or more of the least important features are eliminated from the current feature set. This process is repeated on a new feature subset until the number of remaining features meets the requirements. The input feature subset refers to the subset of feature variables that are most important for predicting a specific Class II water quality parameter after being filtered by the recursive feature elimination method.

[0166] Specifically, each Class II water quality parameter (such as DO, TN, TP, etc.) that needs to be retrieved needs to be dynamically screened individually. Taking dissolved oxygen (DO) retrieval as an example: training samples are collected, each sample including 6 candidate features (Chla, TSM, TURB, SDD, Temp, pH) and corresponding measured dissolved oxygen values. Parameters for the recursive feature elimination method are set, for example, using a random forest as the evaluator, setting the final number of features to be retained or setting an importance threshold (e.g., retaining features with importance > 2%), setting the feature elimination step size to 1, and iteratively eliminating the features with the lowest importance until the relative contribution of all retained features in the evaluator is greater than 2%. The final output will be a subset of features most important for predicting dissolved oxygen, for example, {Temp, pH, Chla}.

[0167] The second inversion model is a machine learning model that can indirectly deduce the second type of water quality parameters, such as a stacking ensemble model, an extreme gradient boosting regression model (XGBoost model), or other regressors.

[0168] For example, continuing from the previous example, for dissolved oxygen (DO), the selected feature subset {Chla, TSM, Temp, pH} is input into the trained limiting gradient boosting regression model, and the model output is the DO concentration inversion value for that region.

[0169] In this embodiment, by employing a recursive feature elimination method combined with a random forest model, the most relevant subset of input features for the second type of water quality parameters is dynamically selected, effectively eliminating redundant and noisy features and achieving precise optimization of the model input. The selected feature subset and the second inversion model are used for inversion, effectively overcoming the core technical bottleneck of the weak direct correlation between the second type of water quality parameters (non-optically active parameters) and remote sensing signals.

[0170] Please see Figure 10 This is a flowchart of a water quality inversion method provided in one embodiment of this application. Figure 10 As shown, the method in this embodiment may include the following steps S4031-S4034, and steps S4031-S4034 may be used as a method for... Figure 8 A refinement of step S403 in the illustrated embodiment.

[0171] S4031, construct base learners based on K-nearest neighbor regression, support vector regression and extreme gradient boosting regression models respectively;

[0172] S4032, Construct a meta-learner based on a preset linear regression model. The prediction function of the meta-learner is shown in the following formula:

[0173]

[0174] Where Y is the second type of water quality parameter predicted by the meta-learner, k is the index value of the base learner, and w k Assign weights, f, to the base learner to the meta-learner. k Let X represent the prediction function of the base learner, and let X be the subset of input features.

[0175] S4033, Construct a stacked ensemble model based on the base learner and the meta learner to obtain the second inversion model;

[0176] S4034, Based on the input feature subset, the second type of water quality parameters are obtained by inversion through the second inversion model. The second type of water quality parameters include dissolved oxygen, permanganate index, chemical oxygen demand, ammonia nitrogen, total nitrogen, and total phosphorus.

[0177] It should be noted that, due to the complex relationship between the second type of water quality parameters and the subset of input features, which may contain multiple patterns, a single model may be insufficient to capture all of them. Therefore, the stacked ensemble learning framework employs multiple base learners, each of which is an independent regression algorithm. (Refer to...) Figure 11 K-Nearest Neighbors (KNN) is an instance-based learning algorithm that predicts the average of the target values ​​of the K nearest neighbor training samples in the feature space of the input sample. Its advantages include no need for training parameters, no assumptions about the data distribution, and the ability to capture local features. Support Vector Regression (SVR) is an algorithm based on the principle of structural risk minimization. Its goal is to find a regression hyperplane such that the deviation of most training samples from the hyperplane is within an acceptable range, while making the hyperplane as flat as possible to improve generalization ability. Its advantages include applicability to high-dimensional data and the ability to handle nonlinear relationships (through kernel functions). Extreme Gradient Boosting Regression (XGBoost) is an efficient gradient boosting decision tree algorithm. It iteratively trains a series of decision trees, each learning the residuals of the predictions from all previous trees, and minimizes the loss function through gradient descent. Its advantages include high prediction accuracy, automatic handling of feature interactions, and robustness to outliers. Each base learner acts as a diverse primary predictor, performing preliminary learning and prediction of the relationship between input features and target parameters from different perspectives, resulting in multiple primary predictions.

[0178] It should be noted that the meta-learner is the second-layer model in the stacked ensemble learning framework. It is used to optimally weight and combine the prediction results of each base learner in the first layer to generate the final, more accurate prediction values ​​of the second type of water quality parameters (predictions of non-optically active parameters).

[0179] The prediction function of the meta-learner is shown in the following formula:

[0180]

[0181] Where Y is the second type of water quality parameter predicted by the meta-learner, k is the index value of the base learner, and w k Assign weights, f, to the meta-learner from the base learner. k Let X represent the prediction function of the base learner, and X be the subset of input features. For example, k=1 represents the K-nearest neighbor regression base learner, k=2 represents the support vector regression base learner, and k=3 represents the limiting gradient boosting regression model base learner.

[0182] Understandably, the heterogeneous models trained in the first two steps are integrated into a unified prediction framework. The advantage of stacked ensemble is that it can automatically learn and utilize the strengths of different base learners through meta-learners, compensating for their shortcomings. The resulting second inversion model possesses the ability of K-nearest neighbor regression to capture local patterns, the ability of support vector regression to handle high-dimensional nonlinearity, and the fitting ability of the limiting gradient boosting regression model. Furthermore, the linear combination of meta-learners avoids the overfitting or bias that may exist in a single model.

[0183] It should be noted that the pre-built second inversion model (stacked ensemble model) is used for actual prediction. The input feature subset is a filtered combination of features for the specific second-type water quality parameter to be predicted. This subset is then input into the trained second inversion model for the corresponding parameter to obtain the predicted value. Further iterative processing of all regions generates a spatial distribution map of various non-optically active parameters for the entire water body. The second-type water quality parameters include dissolved oxygen (DO, reflecting the water body's self-purification capacity) and permanganate index (COD). Mn The indicators include: chemical oxygen demand (COD, reflecting the degree of organic pollution), ammonia nitrogen (NH3-N, reflecting eutrophication and pollution status), total nitrogen (TN, a key indicator of eutrophication), and total phosphorus (TP, a key indicator of eutrophication).

[0184] In this embodiment, three algorithms—K-nearest neighbor regression, support vector regression, and extreme gradient boosting regression—are used to construct diverse base learners. These learners learn the complex relationships between input features and second-type water quality parameters from multiple levels, providing preliminary prediction results. Then, a linear regression model is used as a meta-learner. By weightedly combining the preliminary prediction results, the advantages of different algorithms are fully utilized. Finally, a stacked ensemble strategy is used to combine the base learners and meta-learners into a second inversion model. This model can effectively capture the correlation between optically active parameters, environmental factors, and second-type water quality parameters, improving the inversion accuracy and model generalization ability for key water quality parameters such as dissolved oxygen, chemical oxygen demand, total nitrogen, and total phosphorus.

[0185] It should be understood that although the steps in the flowchart are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order constraint on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the diagram may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0186] Please see Figure 12 One embodiment of this application provides a water quality inversion device. For example... Figure 12 As shown, the water quality inversion device 800 includes:

[0187] The acquisition unit 801 is used to acquire wide-area hyperspectral image data collected by a fixed observation platform, regional multispectral image data collected by a mobile observation platform, and measured water quality parameters collected by ground monitoring points.

[0188] The data processing unit 802 is used to perform correction and spatiotemporal registration processing on wide-area hyperspectral image data and regional multispectral image data, and to perform feature fusion on the registered wide-area hyperspectral image data and the registered regional multispectral image data to generate a fused feature vector.

[0189] The first inversion unit 803 is used to obtain the first type of water quality parameters by inverting the first inversion model based on the fused feature vector.

[0190] The second inversion unit 804 is used to invert the second type of water quality parameters based on the first type of water quality parameters and the measured water quality parameters through the second inversion model.

[0191] The spatial distribution map generation unit 805 is used to invert the spatial distribution map of water quality parameters based on the first type of water quality parameters and the second type of water quality parameters.

[0192] The monitoring unit 806 is used to generate an early warning signal when the measured water quality parameter exceeds a preset threshold, based on the spatial distribution map of water quality parameters.

[0193] Optionally, the data processing unit 802 is further configured to: perform radiometric calibration, atmospheric correction, geometric correction, image stitching, and normalized differential water index extraction on wide-area hyperspectral image data to obtain hyperspectral water reflectance data; perform radiometric calibration, atmospheric correction, geometric correction, image stitching, and normalized differential water index extraction on regional multispectral image data to obtain multispectral reflectance data; and perform spatiotemporal registration of hyperspectral water reflectance data and multispectral reflectance data based on a scale-invariant feature transform algorithm, and perform feature fusion to generate a fused feature vector.

[0194] Optionally, the data processing unit 802 is further configured to: perform spatiotemporal registration of hyperspectral water-leaving reflectance data and multispectral reflectance data based on a scale-invariant feature transform algorithm to obtain registered hyperspectral water-leaving reflectance data and registered multispectral reflectance data; based on the registered hyperspectral water-leaving reflectance data and registered multispectral reflectance data, obtain the reflectance values ​​of the registered hyperspectral water-leaving reflectance data in a preset red band and in a preset near-infrared band, as well as the reflectance values ​​of the registered multispectral reflectance data in the preset red band and in the preset near-infrared band; and calculate a spectral difference compensation factor based on the reflectance values ​​of the registered hyperspectral water-leaving reflectance data in the preset red band and in the preset near-infrared band, as well as the reflectance values ​​of the registered multispectral reflectance data in the preset red band and in the preset near-infrared band, as shown in the following formula:

[0195]

[0196] Where D is the spectral difference compensation factor, R HSI (λ red R represents the reflectance value of the registered hyperspectral water-free reflectance data in the preset red band. HSI (λ NIR R represents the reflectance value of the registered hyperspectral water-leaving reflectance data in the preset near-infrared band. MSI (λ red R represents the reflectance value of the registered multispectral reflectance data in the preset red band. MSI (λ NIR The first part represents the reflectance value of the registered multispectral reflectance data in the preset near-infrared band. Based on the spectral difference compensation factor, the registered hyperspectral water-leaving reflectance data, and the registered multispectral reflectance data, the fused feature vector is obtained, as shown in the following formula:

[0197] F = [HSI, MSI, D]

[0198] Where F is the fused feature vector, HSI is the registered hyperspectral water-free reflectance data, MSI is the registered multispectral reflectance data, and D is the spectral difference compensation factor.

[0199] Optionally, the first inversion unit 803 is further configured to: invert the fused feature vector using a random forest regression model to obtain the chlorophyll a concentration; calculate the reflectance value of a preset band based on the fused feature vector, and invert the total suspended matter, turbidity, and transparency based on the reflectance value of the preset band.

[0200] Optionally, the first inversion unit 803 is also used to: calibrate the inversion coefficients using the least squares method based on measured water quality parameters; and establish inversion models for total suspended solids, turbidity, and transparency based on the inversion coefficients, as shown in the following formulas:

[0201]

[0202] Where Y represents total suspended solids, turbidity, or transparency, and k0 and k i Let λ be the inversion coefficients obtained by calibration using the least squares method. a , λ b The preset band is used; the reflectance value of the preset band is calculated by fusing feature vectors.

[0203] Inversion is performed based on the reflectance value of the preset band, the inversion model of total suspended matter, the inversion model of turbidity, and the inversion model of transparency to obtain total suspended matter, turbidity, and transparency.

[0204] Optionally, the second inversion unit 804 is further configured to: obtain water temperature and pH value from the measured water quality parameters; dynamically filter features based on water temperature, pH value, chlorophyll a concentration, total suspended solids, turbidity, and transparency using a recursive feature elimination method and a random forest regression model to obtain an input feature subset; and invert the second type of water quality parameters based on the input feature subset using a second inversion model.

[0205] Optionally, the second inversion unit 804 is further configured to: construct base learners based on K-nearest neighbor regression, support vector regression, and extreme gradient boosting regression models respectively; and construct a meta-learner based on a preset linear regression model, wherein the prediction function of the meta-learner is shown in the following formula:

[0206]

[0207] Where Y is the second type of water quality parameter predicted by the meta-learner, k is the index value of the base learner, and w k Assign weights to the base learners to the meta-learner, f kLet X represent the prediction function of the base learner, and let X be the subset of input features. A stacked ensemble model is constructed based on the base learner and the meta-learner to obtain the second inversion model. Based on the subset of input features, the second type of water quality parameters are obtained by inversion through the second inversion model. The second type of water quality parameters include dissolved oxygen, permanganate index, chemical oxygen demand, ammonia nitrogen, total nitrogen, and total phosphorus.

[0208] Specific limitations regarding the aforementioned water quality inversion device 800 can be found in the limitations of the water quality inversion method described above, and will not be repeated here. Each module in the aforementioned water quality inversion device 800 can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0209] In one embodiment, a computer device is provided, the internal structure of which can be as follows: Figure 13 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the water quality inversion method described above. It includes: a memory and a processor; the memory stores a computer program; and the processor executes the computer program to implement any step of the water quality inversion method described above.

[0210] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, can perform any step of the water quality inversion method described above.

[0211] Those skilled in the art will understand that embodiments of this application can provide methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0212] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0213] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0214] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0215] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0216] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A water quality inversion method, characterized in that, include: Acquire wide-area hyperspectral image data collected by fixed observation platforms, regional multispectral image data collected by mobile observation platforms, and measured water quality parameters collected by ground monitoring points; The wide-area hyperspectral image data and the regional multispectral image data are corrected and spatiotemporally registered, and the registered wide-area hyperspectral image data and the registered regional multispectral image data are feature-fused to generate a fused feature vector. Based on the fused feature vector, the first type of water quality parameters are obtained by inversion through the first inversion model; Based on the first type of water quality parameters and the measured water quality parameters, the second type of water quality parameters are obtained by inversion through the second inversion model; A spatial distribution map of water quality parameters is obtained by inverting the first type of water quality parameters and the second type of water quality parameters. Based on the spatial distribution map of the water quality parameters, when the measured water quality parameters are greater than a preset threshold, an early warning signal is generated; The process of correcting and spatiotemporally registering the wide-area hyperspectral image data and the regional multispectral image data, and then fusing the registered wide-area hyperspectral image data with the registered regional multispectral image data to generate a fused feature vector includes: Radiometric calibration, atmospheric correction, geometric correction, image stitching, and normalized differential water index extraction were performed on the wide-area hyperspectral image data to obtain hyperspectral water reflectance data. Radiometric calibration, atmospheric correction, geometric correction, image stitching, and normalized differential water index extraction were performed on the multispectral image data of the region to obtain multispectral reflectance data. Based on the scale-invariant feature transformation algorithm, the hyperspectral water-free reflectance data and the multispectral reflectance data are spatiotemporally registered and feature fused to generate a fused feature vector. The scale-invariant feature transformation algorithm performs spatiotemporal registration on the hyperspectral water-leaving reflectance data and the multispectral reflectance data, and performs feature fusion to generate a fused feature vector, including: Based on the scale-invariant feature transform algorithm, the hyperspectral water-free reflectance data and the multispectral reflectance data are spatiotemporally registered to obtain registered hyperspectral water-free reflectance data and registered multispectral reflectance data. Based on the registered hyperspectral water-free reflectance data and the registered multispectral reflectance data, obtain the reflectance values ​​of the registered hyperspectral water-free reflectance data in a preset red band and in a preset near-infrared band, as well as the reflectance values ​​of the registered multispectral reflectance data in a preset red band and in a preset near-infrared band. Based on the registered hyperspectral water-leaving reflectance data in the preset red band and the preset near-infrared band, and the registered multispectral reflectance data in the preset red band and the preset near-infrared band, the spectral difference compensation factor is calculated, as shown in the following formula: ； Where D is the spectral difference compensation factor. The reflectance value of the registered hyperspectral water-free reflectance data in the preset red light band. The reflectance value of the registered hyperspectral water-leaving reflectance data in the preset near-infrared band. The reflectance value of the registered multispectral reflectance data in the preset red light band. The reflectance value of the registered multispectral reflectance data in the preset near-infrared band; Based on the spectral difference compensation factor, the registered hyperspectral water-free reflectance data, and the registered multispectral reflectance data, the fused feature vector is obtained, as shown in the following formula: F=[HSI,MSI,D] Where F is the fused feature vector, HSI is the registered hyperspectral water-free reflectance data, MSI is the registered multispectral reflectance data, and D is the spectral difference compensation factor.

2. The water quality inversion method according to claim 1, characterized in that, The first category of water quality parameters includes chlorophyll a concentration, total suspended solids, turbidity, and transparency; the first category of water quality parameters is obtained by inversion using a first inversion model based on the fused feature vector, including: The chlorophyll a concentration is obtained by inverting the fused feature vector using a random forest regression model. The reflectance value of the preset band is calculated based on the fused feature vector, and the total suspended matter, the turbidity, and the transparency are obtained by inversion based on the reflectance value of the preset band.

3. The water quality inversion method according to claim 2, characterized in that, The step of calculating the reflectance value of a preset band based on the fused feature vector, and then performing inversion based on the reflectance value of the preset band to obtain the total suspended matter, the turbidity, and the transparency includes: Based on the measured water quality parameters, the inversion coefficients were obtained by calibration using the least squares method. Based on the inversion coefficients, inversion models are established for the total suspended matter, the turbidity, and the transparency, as shown in the following formulas: ； Where Y represents total suspended solids, turbidity, or transparency. k i Let λ be the inversion coefficients obtained by calibration using the least squares method. a , λ b Preset band; The reflectance value of the preset band is calculated using the fused feature vector; The total suspended matter, turbidity, and transparency are obtained by inversion based on the reflectance value of the preset band, the inversion model of the total suspended matter, the inversion model of the turbidity, and the inversion model of the transparency.

4. The water quality inversion method according to claim 2, characterized in that, The step of obtaining the second type of water quality parameters by inversion using a second inversion model based on the first type of water quality parameters and the measured water quality parameters further includes: Obtain the water temperature and pH value from the measured water quality parameters; Based on the water temperature, pH value, chlorophyll a concentration, total suspended matter, turbidity, and transparency, a subset of input features is obtained by dynamic feature selection using recursive feature elimination and random forest regression models. Based on the input feature subset, the second type of water quality parameters are obtained by inversion using the second inversion model.

5. The water quality inversion method according to claim 4, characterized in that, The step of obtaining the second type of water quality parameters by inverting the second inversion model based on the input feature subset includes: Base learners are constructed based on K-nearest neighbor regression, support vector regression, and extreme gradient boosting regression models, respectively. A meta-learner is constructed based on a pre-defined linear regression model, and the prediction function of the meta-learner is specifically shown in the following formula: ； Where Y is the second type of water quality parameter predicted by the meta-learner, and k is the index value of the base learner. Assign weights to the base learner to the meta learner. Let X represent the prediction function of the base learner, and let X be the subset of input features. A stacked ensemble model is constructed based on the base learner and the meta learner to obtain the second inversion model; Based on the input feature subset, the second type of water quality parameters are obtained by inversion through the second inversion model. The second type of water quality parameters include dissolved oxygen, permanganate index, chemical oxygen demand, ammonia nitrogen, total nitrogen, and total phosphorus.

6. A water quality inversion device, characterized in that, include: The acquisition unit is used to acquire wide-area hyperspectral image data collected by fixed observation platforms, regional multispectral image data collected by mobile observation platforms, and measured water quality parameters collected by ground monitoring points. The data processing unit is used to perform correction and spatiotemporal registration processing on the wide-area hyperspectral image data and the regional multispectral image data, and to perform feature fusion on the registered wide-area hyperspectral image data and the registered regional multispectral image data to generate a fused feature vector. The first inversion unit is used to invert the first type of water quality parameters based on the fused feature vector and through the first inversion model. The second inversion unit is used to invert the second type of water quality parameters based on the first type of water quality parameters and the measured water quality parameters through the second inversion model; The spatial distribution map generation unit is used to invert the spatial distribution map of water quality parameters based on the first type of water quality parameters and the second type of water quality parameters; The monitoring unit is used to generate an early warning signal when the measured water quality parameter is greater than a preset threshold, based on the spatial distribution map of the water quality parameter. The data processing unit is further configured to: Radiometric calibration, atmospheric correction, geometric correction, image stitching, and normalized differential water index extraction were performed on the wide-area hyperspectral image data to obtain hyperspectral water reflectance data. Radiometric calibration, atmospheric correction, geometric correction, image stitching, and normalized differential water index extraction were performed on the multispectral image data of the region to obtain multispectral reflectance data. Based on the scale-invariant feature transformation algorithm, the hyperspectral water-free reflectance data and the multispectral reflectance data are spatiotemporally registered and feature fused to generate a fused feature vector. The data processing unit is further configured to: Based on the scale-invariant feature transform algorithm, the hyperspectral water-free reflectance data and the multispectral reflectance data are spatiotemporally registered to obtain registered hyperspectral water-free reflectance data and registered multispectral reflectance data. Based on the registered hyperspectral water-free reflectance data and the registered multispectral reflectance data, obtain the reflectance values ​​of the registered hyperspectral water-free reflectance data in a preset red band and in a preset near-infrared band, as well as the reflectance values ​​of the registered multispectral reflectance data in a preset red band and in a preset near-infrared band. Based on the registered hyperspectral water-leaving reflectance data in the preset red band and the preset near-infrared band, and the registered multispectral reflectance data in the preset red band and the preset near-infrared band, the spectral difference compensation factor is calculated, as shown in the following formula: ； Where D is the spectral difference compensation factor. The reflectance value of the registered hyperspectral water-free reflectance data in the preset red light band. The reflectance value of the registered hyperspectral water-leaving reflectance data in the preset near-infrared band. The reflectance value of the registered multispectral reflectance data in the preset red light band. The reflectance value of the registered multispectral reflectance data in the preset near-infrared band; Based on the spectral difference compensation factor, the registered hyperspectral water-free reflectance data, and the registered multispectral reflectance data, the fused feature vector is obtained, as shown in the following formula: F=[HSI,MSI,D] Where F is the fused feature vector, HSI is the registered hyperspectral water-free reflectance data, MSI is the registered multispectral reflectance data, and D is the spectral difference compensation factor.

7. A computer device, comprising: The method includes a memory and a processor, the memory storing a computer program, characterized in that the processor executes the computer program to implement the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images