A monitoring and intelligent identification method and system for small and micro black and odorous water bodies

Abstract

The application discloses a kind of monitoring and intelligent identification method and system for small and micro black and odorous water body, belong to black and odorous water body monitoring technical field. Including star ground data matching stage, optimal feature mining stage, AI spectrum super-resolution reconstruction stage and multi-index business grading identification stage. Utilize three kinds of data sources such as hyperspectral satellite, unmanned aerial vehicle, ground measurement collaborative processing, can realize large-scale water body screening and small-scale black and odorous water body fine identification simultaneously, overcome the inherent limitation of single data source. First rely on satellite and ground measurement data to mine the core spectral features sensitive to water quality, then migrate the part of spectral prior to the high spatial resolution unmanned aerial vehicle observation scene, transfer spectral information to centimeter level spatial scale by means of super-resolution reconstruction, finally combine business standard to complete black and odorous water body pollution grade determination, realize the combination of wide-area satellite screening capability and local fine identification capability.

Description

Technical Field

[0001] This invention relates to a monitoring and intelligent identification method and system for small-scale black and odorous water bodies, belonging to the field of black and odorous water body monitoring technology. Background Technology

[0002] Small-scale black and odorous water bodies are a key issue in my country's current efforts to improve the living environment and manage water ecology. Their formation is typically related to factors such as direct discharge of domestic sewage, pollution from livestock and poultry farming, agricultural non-point source pollution, insufficient water flow, and anaerobic decomposition of sediment. Black and odorous water bodies not only exhibit abnormal color, decreased transparency, and a pronounced odor, but are also accompanied by typical water environment deterioration phenomena such as increased ammonia nitrogen, decreased dissolved oxygen, and enhanced organic pollution. Currently, my country has issued relevant normative documents such as the "Guidelines for the Treatment of Urban Black and Odorous Water Bodies" and the "Guidelines for the Treatment of Rural Black and Odorous Water Bodies (Trial Implementation)," which set forth clear requirements for the identification, classification, and treatment of black and odorous water bodies. Among these, transparency (SD), ammonia nitrogen (NH3-N), and dissolved oxygen (DO) have become important criteria for judgment.

[0003] Currently, the traditional methods for monitoring black and odorous water bodies mainly rely on manual on-site inspections combined with point sampling and testing. On-site sampling and testing can obtain relatively accurate water quality measurement data, but it requires a large manpower investment, has limited monitoring coverage, and has a long data update cycle, making it difficult to complete the routine and dynamic monitoring of water bodies in a large area. Remote sensing technology has therefore gradually become an important technical means for monitoring black and odorous water bodies. The main technologies currently available are as follows, but certain problems still exist: 1. Black and odorous water body identification methods based on remote sensing indices. Hyperspectral remote sensing has continuous narrow-band spectral information, which can better reflect the physicochemical characteristics of water bodies and has a significant advantage in water quality parameter inversion; UAV remote sensing has centimeter-level spatial resolution and flexible deployment capabilities, suitable for localized, fine-grained inspections. However, most existing technical solutions use only satellite remote sensing or UAV remote sensing as a single data source for water body monitoring. Identification methods based on remote sensing indices rely on empirically set thresholds to determine the black and odorous nature of water bodies. While computationally efficient, they have poor adaptability in mixed ground cover scenarios and cannot achieve quantitative inversion of key water quality physicochemical parameters such as ammonia nitrogen and dissolved oxygen. 2. Water quality inversion methods based on hyperspectral remote sensing. The technology of water quality inversion using hyperspectral data generally uses all bands in model construction. However, the sheer number of hyperspectral bands and the strong collinearity between them lead to data dimensional redundancy, making the model prone to feature importance drift and poor generalization across time periods. Furthermore, hyperspectral satellites are limited by hardware conditions, resulting in limited spatial resolution and difficulty in identifying widely distributed small-scale water bodies such as narrow ditches and small ponds. 3. Water environment monitoring methods based on UAV remote sensing. Ordinary RGB sensors on UAVs can only output red, green, and blue channel images, lacking continuous spectral information. Even with deep learning networks to extract water body areas, the lack of physical constraints easily leads to spectral distortion and makes it impossible to quantitatively invert core water quality indicators. 4. Deep learning-based remote sensing super-resolution methods. Most existing remote sensing image super-resolution algorithms aim to improve image visual clarity, belonging to purely data-driven black-box models. They do not incorporate the physical constraints of the water body spectrum itself, making the reconstruction results prone to spectral feature distortion, directly reducing the reliability of subsequent water quality inversion results. At the same time, existing research mainly focuses on inversion or water body classification of single water quality parameters, lacking multi-indicator collaborative identification logic that conforms to industry standards. It has not established a complete technical link for the collaborative linkage of three types of data: hyperspectral satellite, UAV, and ground measurement. It cannot simultaneously meet the practical needs of rapid screening of large-scale water bodies and fine identification of small-scale water bodies, and it is difficult to directly support the automated and operational supervision of small and micro black and odorous water bodies.

[0004] In summary, while existing technologies have made some progress in hyperspectral remote sensing, water quality inversion, UAV monitoring, and super-resolution reconstruction, a method for intelligent monitoring and hierarchical identification of small and micro black and odorous water bodies that can integrate the advantages of hyperspectral spectroscopy, the high spatial resolution of UAVs, and the physical constraint AI reconstruction mechanism is still lacking. Therefore, there is an urgent need for an intelligent monitoring and identification technology for small and micro black and odorous water bodies that can balance wide-area monitoring with localized fine-grained identification. Summary of the Invention

[0005] The purpose of this invention is to address the aforementioned problems by providing a method and system for monitoring and intelligent identification of small-scale black and odorous water bodies. This system utilizes the spectral characteristics of wide-area hyperspectral satellite data as the baseline spectral prior knowledge for local microscopic observations. Through a SHAP interpretive mechanism with stability constraints, it extracts low-dimensional targeted optimal features that have statistically sensitive responses to target water quality indicators. Furthermore, it guides the construction of a super-resolution reconstruction network embedded with physical constraints of spectral gradients using local UAV true-color (RGB) data, achieving spatial morphological mapping and abundance enhancement of optimal features at the microscopic scale. Finally, combined with ground-based synchronous measurement data correction, and through multi-indicator collaborative mapping and a fixed threshold discrimination matrix, it achieves refined classification and identification of small-scale black and odorous water bodies.

[0006] The technical solution adopted in this invention is as follows: A monitoring and intelligent identification method for small-scale black and odorous water bodies includes the following steps: S1. Satellite-Ground Data Matching Stage: Data preprocessing is performed on hyperspectral satellite data to obtain processed satellite layers. Data matching is performed between the processed satellite layers and the measured point data to complete spatial grid coordinate alignment and pixel purification. S2, Optimal Feature Mining Stage: After completing data matching, the data is input into the primary inversion model of water quality parameters at the space-to-ground scale to establish the mapping relationship between spectral features and water quality indicators; and the SHAP evaluation with stability constraint mechanism is introduced to perform dimensionality reduction screening by quantifying the marginal contribution of each spectral feature, and output the stable and reproducible optimal features. S3, AI Spectral Super-Reconstruction Stage: The UAV data and the optimal features output from S2 are used as bidirectional input variables to jointly perform AI spectral super-reconstruction and output a local optimal feature distribution map. S4, Multi-indicator Business Classification and Identification Stage: The local optimal feature distribution map output from S3 is used as input to inject and construct a UAV-scale water quality inversion model to invert a high spatial resolution water quality parameter distribution map; the water quality parameter distribution map is then used for multi-indicator collaborative water body business classification mapping, and a classification judgment is made through a fixed threshold discrimination matrix, with independent business classification results output for each.

[0007] This scheme completes hyperspectral satellite data preprocessing and data matching with measured data points in step S1, achieving spatial alignment between satellite remote sensing spectral information and ground-measured water quality information. Pixel purification reduces spectral interference from non-water features along the shore, constructing a reliable satellite-ground joint sample foundation and providing clean and effective input data for subsequent model training. Based on this, step S2 establishes a mapping relationship between spectral features and water quality indicators using a primary satellite-ground scale water quality parameter inversion model. Combined with SHAP evaluation with stability constraints, spectral feature dimensionality reduction and screening are performed. This allows for the extraction of optimal features with stable responses to water quality indicators from high-dimensional redundant hyperspectral bands, avoiding band collinearity and model output fluctuations caused by full-band modeling, compressing the solution space of subsequent models, and providing physically statistically significant spectroscopic prior knowledge for the super-resolution reconstruction stage. Step S3 inputs both UAV data and optimal features into the AI ​​spectral super-resolution reconstruction process. This combines the spectral prior information extracted from satellites with the high spatial resolution information possessed by the UAV, compensating for the lack of continuous water quality-sensitive spectral information in the original UAV imagery. This generates a local optimal feature distribution map that combines spatial detail with effective spectral characterization. Step S4 uses this local optimal feature distribution map to build a UAV-scale water quality inversion model, obtaining high spatial resolution water quality parameter distribution results. Then, through multi-index collaborative water body operational classification mapping and a fixed threshold discrimination matrix, classification is completed, directly outputting water body classification results that meet operational regulatory needs. This establishes a complete link from remote sensing data processing to automated identification of water pollution levels.

[0008] This solution utilizes three data sources—hyperspectral satellites, UAVs, and ground-based measurements—for collaborative processing, enabling simultaneous large-scale water body screening and refined identification of small-scale black and odorous water bodies, overcoming the inherent limitations of single data sources. First, it mines core spectral features sensitive to water quality based on satellite and ground-based data. Then, it transfers these spectral priors to high spatial resolution UAV observation scenarios. Super-resolution reconstruction transfers spectral information to a centimeter-level spatial scale. Finally, it combines operational standards to determine the water pollution level, achieving a combination of wide-area satellite screening capabilities and localized refined identification capabilities. This solution overcomes the inherent limitations of insufficient spatial resolution in hyperspectral satellites and the lack of water-sensitive spectral information in UAV imagery. The feature selection process uses stability constraints to ensure the reusability of selected spectral features across time phases, avoiding the insufficient generalization ability of traditional full-band modeling. The super-resolution reconstruction process uses optimal features obtained from satellites as prior guidance, changing the traditional data-driven model that only focuses on visual effect optimization. This allows the reconstructed spatial map to carry effective spectral information related to water quality indicators, which is then directly connected to the discrimination matrix of operational standards, transforming remote sensing inversion results into directly applicable regulatory classification conclusions.

[0009] Compared to other remote sensing index-based identification techniques, this scheme leverages satellite-to-ground matching sample mining to obtain optimal features statistically correlated with water quality indicators. It eliminates reliance on simple water body indices for discrimination, enabling the acquisition of spatial distribution results for multiple key water quality parameters and better adapting to complex underlying surface environments. Compared to hyperspectral remote sensing-based water quality inversion techniques, this scheme introduces stability-constrained feature selection logic on top of the satellite-to-ground scale primary inversion model. It extracts core features of stable responses from high-dimensional spectra, reducing interference from redundant bands and improving the model's robustness and cross-temporal adaptability. The selected spectral priors are then transferred to UAV observation scenarios, mapping water quality-related spectral information to a centimeter-level spatial scale, compensating for the limitations of satellite spatial resolution and enabling the monitoring and identification of small-scale black and odorous water bodies. Compared to monitoring methods relying on UAV remote sensing, this solution does not directly use raw UAV imagery for water quality calculations. Instead, it feeds the optimal features mined from satellite data as prior information, along with UAV data, into the AI ​​spectral super-resolution reconstruction stage. This supplements the high-resolution imagery with water-sensitive spectral information, providing the UAV observation data with the fundamental conditions for quantitative inversion of transparency, ammonia nitrogen, and dissolved oxygen, fully leveraging the high spatial resolution advantage of UAVs. Compared to deep learning-based remote sensing super-resolution methods, this solution uses the optimal features output from satellite as prior guidance for the super-resolution process. This ensures that the reconstruction process serves the spatial mapping of water-sensitive features, rather than simply optimizing visual imaging effects, reducing non-physical spectral distortions and ensuring the reliability of subsequent water quality parameter inversion results. Furthermore, after obtaining high-resolution multi-water quality parameter distribution maps, this solution uses a fixed threshold discrimination matrix conforming to industry standards to conduct multi-indicator collaborative classification, automatically completing the classification output of normal water bodies into mildly and severely polluted water. This forms a complete automated closed loop from data preprocessing, feature mining, super-resolution reconstruction, and water quality inversion to business classification and identification. The output results can directly serve the routine supervision of small and micro polluted water bodies.

[0010] Alternatively, the data preprocessing in S1 may include: Radiometric calibration was performed on the hyperspectral satellite data to eliminate sensor system errors; the FLAASH model was used to perform atmospheric correction on the radiometrically calibrated data to obtain the true remote sensing reflectance of the water surface after eliminating atmospheric interference.

[0011] Radiometric calibration of hyperspectral satellite data can eliminate systematic errors introduced by the sensor itself, transforming the raw satellite observation signal into a radiometric quantity with physical meaning. Subsequent atmospheric correction using the FLAASH model can remove signal interference from atmospheric scattering and absorption, yielding remotely sensed water surface reflectance that closely approximates the true state of the water body. The spectral data obtained through this process accurately reflects the spectral response characteristics of the water body itself, reducing spurious signal interference from the sensor and the atmospheric environment.

[0012] Alternatively, the data matching in S1 may include: A coarse water mask was extracted from the satellite layer using the normalized differential water index. For pixels at the mask boundary, a linear spectral unmixing model was used to perform abundance inversion with water, vegetation, and soil as endmembers. Only pixels with pure water abundance greater than the target value were retained to complete pixel purification and obtain the satellite pixel spectral reflectance curve. The purified satellite pixel spectral reflectance curve was then spatially aligned with the ground-synchronized measured point data to construct a satellite-ground integrated primary training sample library.

[0013] By utilizing the normalized differential water index, the approximate distribution range of water bodies can be quickly delineated, generating a coarse water body mask. For pixels with mixed land cover at the mask boundaries, a linear spectral unmixing model is used to select water body vegetation and soil as endmembers for abundance inversion. Only pixels with pure water abundance reaching the target value are retained, which can filter out spectral pollution caused by land cover such as shoreline vegetation and soil, resulting in a purer satellite pixel spectral reflectance curve. The purified satellite pixel spectral reflectance curve is then spatially aligned with the ground-based synchronously acquired measured data, achieving a one-to-one spatial correspondence between remote sensing spectral information and on-site water quality monitoring data. The resulting integrated satellite-ground primary training sample library reduces the proportion of anomalous samples introduced by mixed pixels, minimizes interference from non-water body spectral signals on the model learning process, and improves the overall data reliability of the sample library.

[0014] Alternatively, the methods for constructing the primary inversion model of water quality parameters at the terrestrial and space scales in S2 include: Using purified hyperspectral satellite remote sensing reflectance in each band as input independent variables and ground-measured water quality indicators as dependent variables, an ensemble learning primary inversion model based on LightGBM is constructed to fit the nonlinear mapping pathway between high-dimensional spectral space and water quality indicators.

[0015] The characteristics of ensemble learning models can take into account the information contribution of multiple sets of spectral input variables, have strong noise resistance, adapt to the non-monotonic variation of water body spectra with water quality changes, and rely on a high-quality satellite-ground integrated primary training sample library for training. It can establish a preliminary mapping path for the conversion of spectral signals into actual water quality indicators, and provide an interpretable model carrier for the subsequent SHAP evaluation to analyze the marginal contribution of each spectral feature to water quality indicators.

[0016] Alternatively, the SHAP evaluation that introduces a stability constraint mechanism in S2 specifically includes: A1. Perform M rounds of Bootstrap self-sampling on the historical sample dataset, which contains T sample time phases; A2. In each round of sampling, the SHAP game theory framework is used to perform inverse analysis of the model, calculating the Shapley marginal contribution value of each spectral feature. The calculation expression is as follows: In the formula, Let F be the marginal contribution value of spectral feature i, F be the full-spectrum feature set, and S be the feature subset excluding feature i. This is the model's predicted output on the feature subset S; A3. Statistical analysis of the absolute values ​​of sapride for each spectral characteristic under M rounds of sampling. The variance is used to retain only the high-frequency intersection features whose variance is less than a set threshold and which rank in the top N of the multi-temporal sample sets, and these features are established as the optimal features.

[0017] Multiple rounds of Bootstrap self-sampling are conducted on historical sample datasets containing multiple sample phases to generate multiple training subsets with sample differences, simulating the model's output performance under different sample conditions. After each round of sampling, the SHAP game theory framework is used to perform inverse analysis on the trained primary inversion model, calculating the Shapley marginal contribution value corresponding to each spectral feature. This quantifies the contribution of each spectral feature to the water quality index prediction results, objectively reflecting the response degree of different bands to water quality changes. The variance of the absolute value of each spectral feature under all sampling rounds is calculated. Combined with the feature contribution ranking under multiple time phase sample conditions, high-frequency intersection features are screened. This eliminates spectral bands that only have high contributions in some sample scenarios and whose contribution fluctuates drastically with sample perturbation, retaining only features with stable contribution performance across sampling rounds and multiple time phase conditions as the optimal features.

[0018] Alternatively, the AI ​​spectral super-resolution reconstruction in S3 may specifically include: B1. Construct a convolutional neural network model based on an encoder-decoder architecture; B2. The UAV data and the optimal features output by S2 are concatenated at the front end of the encoder of the convolutional neural network model to achieve multimodal feature concatenation. B3. During the backpropagation stage of the convolutional neural network model training, a loss function embedding physical constraints of spectral gradients is designed to constrain the reconstruction process. The loss function is: In the formula, This is the mean square error loss term. For spatial continuity constraints, For the morphological consistency constraint term of the spectroscopic gradient, and This is the weighting adjustment coefficient; B4. After training and iterating the convolutional neural network model, output the optimal feature space distribution map with high spatial resolution.

[0019] A convolutional neural network with an encoder-decoder architecture is constructed to perform feature extraction and upsampling reconstruction of multimodal input information. By concatenating the best features from UAV data and S2 output at the encoder front end, multimodal feature cascading is achieved. This simultaneously incorporates high spatial detail information from the UAV and stable spectral priors with water quality response capabilities obtained from satellite imagery. The network learns both spatial texture information and water-sensitive spectral features, avoiding unguided blind reconstruction based solely on UAV imagery. A composite loss function with multiple components is introduced during the model's backpropagation stage to constrain unconstrained random fitting during AI reconstruction, mitigating non-physical numerical jumps in the spectral dimension and neighborhood of reconstructed features. The mean squared error loss term ensures a good numerical fit between the reconstructed output and the reference ground truth. The spatial continuity constraint maintains a smooth transition in the spatial distribution within the water body and suppresses pseudo-texture noise generated during super-resolution. The spectral gradient morphology consistency constraint constrains the spectral change trend after reconstruction, ensuring the morphology of the spectral curve remains consistent with the original satellite optimal features. Weight adjustment coefficients flexibly adjust their respective contribution proportions, achieving a balance between spatial detail restoration and spectral physical accuracy. After multiple rounds of training iterations to achieve network parameter convergence, the optimal feature space distribution map with high spatial resolution is output, which can obtain a feature map that has both centimeter-level spatial details and reliable water mass spectrometry meaning.

[0020] Alternatively, the spatial continuity constraint term The expression is: In the formula, A is the optimal feature; i represents the row index of the image pixel; j represents the column index of the image pixel; ∑i,j represents the summation of the spectral gradient differences of each pixel position in the reconstructed region; ∥⋅∥ represents the norm used to measure the difference between the UAV reconstructed spectral gradient and the satellite prior spectral gradient. The spectral gradient morphological consistency constraint term The expression is: In the formula, A λa and A λbλa and λb represent the eigenvalues ​​corresponding to the adjacent or neighboring optimal spectral features with wavelengths λa and λb, respectively; λa and λb represent the center wavelengths of the bands corresponding to the two optimal spectral features, respectively; the subscript UAV indicates the high spatial resolution optimal feature obtained by AI spectral super-resolution reconstruction of UAV imagery; the subscript Sat_Prior indicates the prior of the corresponding optimal spectral feature determined by hyperspectral satellite data; i represents the row index of the image pixel; j represents the column index of the image pixel; ∑i,j represents the accumulation of the spectral gradient differences at each pixel position within the reconstruction region; ∥⋅∥ represents the norm used to measure the difference between the UAV reconstructed spectral gradient and the satellite prior spectral gradient.

[0021] The spatial continuity constraint term traverses all pixel positions in the image, calculates the feature difference norm between each pixel and its horizontal and vertical adjacent pixels, and performs a summation operation. This can constrain the spatial gradient of the super-resolution reconstruction output feature map, guide the network to output a feature distribution result with a gradual change in the homogeneous water region, suppress the checkerboard artifacts and random noise that are easily generated in the super-resolution process of convolutional networks, and reduce false feature signals that are not real abrupt changes.

[0022] The spectral gradient morphology consistency constraint term constrains the first-order trend of the reconstructed spectral curve, ensuring that the slope of the inter-band variation of the local features obtained by UAV-based super-resolution closely approximates the spectral line morphology established a priori by the satellite. When the optimal feature includes water quality diagnostic bands, this constraint can lock the relative slope boundary between adjacent bands, suppressing spectral morphology distortions that are prone to occur in the deep learning super-resolution process, and avoiding the generation of false spectral curves that violate the physical laws of water bodies.

[0023] Alternatively, the method for constructing the UAV-scale water quality inversion model in S4 includes: The local centimeter-level optimal feature distribution map generated by AI spectral super-resolution reconstruction is used as input to map and output spatial thematic maps of three water quality indicators, namely transparency, ammonia nitrogen, and dissolved oxygen, at centimeter resolution. The spatial thematic maps of the three water quality indicators are then mapped to a fixed threshold discrimination matrix in the standard administrative business specifications.

[0024] Quantitative inversion is performed using high-fidelity optimal features obtained through dual physical constraints. Since the optimal features have filtered out some noise interference, and the spectral response patterns of water bodies mined from satellites are inherited, the prior knowledge of water quality spectra from large-scale satellites can be transferred to the centimeter-level UAV scale. The centimeter-level resolution can capture the spatial heterogeneity of water quality within small water bodies such as ditches and ponds, and can identify water quality differences in local areas of water bodies. Simultaneously, centimeter-level spatial thematic mapping of three key indicators—transparency, ammonia nitrogen, and dissolved oxygen—is completed, which can capture water quality abrupt changes in small areas such as ditch corners, bank boundaries, and the vicinity of local sewage outlets. Mapping the thematic maps to a fixed threshold discrimination matrix of administrative business specifications can complete the alignment of coordinate system one, hierarchical standards, and business format adaptation, allowing the output water quality thematic results to be directly connected to water environment supervision business.

[0025] Alternatively, the determination rule for the fixed threshold discrimination matrix can be: In the fixed threshold discrimination matrix, transparency, ammonia nitrogen, and dissolved oxygen each have a preset threshold for normal water body determination, a threshold for mild black and odorous water body determination, and a threshold for severe black and odorous water body determination. The fixed threshold discrimination matrix performs a pixel-by-pixel single-factor veto cascade scan. If any water quality indicator of any pixel triggers its corresponding pollution determination threshold, the pixel is determined to be of the corresponding level. If none of the indicators trigger the pollution determination threshold, the pixel is determined to be normal water body.

[0026] It aligns with the business evaluation logic of black and odorous water body assessment, avoids the problem of pollution information being masked by multi-indicator weighted averaging, and avoids the misjudgment of a single parameter exceeding the standard but the overall score being qualified. At the same time, relying on pixel-by-pixel scanning, it can accurately locate local abnormal points such as local sewage outlets and corner pollution patches in small water bodies. The judgment logic is simple, transparent, and highly reproducible, making it easy to match with water environment supervision standards. The output results can directly support actual supervision work such as locating black and odorous patches, delineating the scope of pollution, and tracing the source of problematic water bodies.

[0027] A monitoring and intelligent identification system for small and micro black and odorous water bodies includes: The data preprocessing module standardizes the hyperspectral satellite data and outputs the processed satellite layers. The data matching module aligns the processed satellite layer with the measured point data using spatial grid coordinates and performs pixel purification to construct a primary training sample library integrating satellite and ground. The primary inversion model for water quality parameters at the space-to-ground scale uses data from the space-to-ground integrated primary training sample library as input to establish the mapping relationship between spectral features and water quality indicators, and outputs the mapping model between spectral features and water quality indicators. The SHAP evaluation module performs dimensionality reduction and screening on the mapping relationship output by the primary inversion model of water quality parameters at the terrestrial and space scales by quantifying the marginal contribution of each spectral feature, and outputs the optimal features that are stable and reproducible. The AI ​​spectral super-resolution reconstruction module takes UAV data and optimal features as bidirectional input variables, and outputs a local optimal feature distribution map through super-resolution reconstruction processing with embedded spectral gradient physical constraints. The UAV-scale water quality inversion model uses the local optimal feature distribution map as input to invert a high spatial resolution water quality parameter distribution map. The multi-indicator collaborative water body business classification mapping module maps the water quality parameter distribution map to a fixed threshold discrimination matrix of standard administrative business specifications; The fixed threshold discrimination matrix module performs multi-level logic gate judgments on the mapped water quality parameter distribution map, and outputs different business classification results respectively.

[0028] By fusing hyperspectral satellite data with centimeter-level UAV imagery and combining it with ground-based measured data, and relying on SHAP interpretability analysis to complete feature selection, the stable and optimal features mined from satellites are transferred to the centimeter-level scale of UAVs. At the same time, physical constraints of spectral gradients are introduced in the super-resolution stage to suppress the spectral distortion problem of AI models, solving the industry shortcomings of insufficient large-scale satellite resolution in traditional remote sensing monitoring, lack of large-scale prior constraints in UAV monitoring, and weak physical interpretability of pure AI model results. The system backend connects to the administrative business dimension mapping and fixed threshold discrimination matrix to realize the full-process automation from raw remote sensing data processing, model training, high-resolution mapping to black and odorous water body classification identification. The output results are directly adapted to the business needs of water environment supervision, taking into account the large-area satellite survey capabilities and the local fine-grained investigation capabilities of UAVs, realizing the integrated air-ground intelligent identification of water bodies.

[0029] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. The present invention provides a method and system for monitoring and intelligent identification of small-scale black and odorous water bodies. It achieves multi-source collaborative fusion of hyperspectral satellite data, UAV centimeter-level imagery, and ground-based measured data, realizing multi-scale collaborative monitoring that combines macroscopic survey coverage with microscopic local fine identification. This overcomes the inherent shortcomings of single data sources in terms of spatial resolution or spectroscopic information, comprehensively acquiring the spectral and spatial detail information of small-scale water bodies, effectively improving the overall completeness of monitoring.

[0030] 2. The present invention provides a monitoring and intelligent identification method and system for small and micro black and odorous water bodies. It introduces a stability-constrained SHAP feature evaluation mechanism to process high-dimensional hyperspectral data, and selects the optimal features that have a stable response to water quality indicators. This can reduce the adverse effects of redundant bands and band collinearity, improve the interpretability of the features themselves, strengthen the generalization ability of the model under cross-temporal conditions, and avoid process disturbances caused by feature importance drift.

[0031] 3. The monitoring and intelligent identification method and system for small-scale black and odorous water bodies provided by this invention embeds physical constraints of spectral gradients into the AI ​​super-resolution reconstruction process. This allows the reconstruction results to maintain both high spatial resolution and the continuity and rationality of spectral changes, suppressing spectral distortion and improving the physical reliability of the output features. The high-resolution features reconstructed based on physical constraints, combined with the high spatial observation capabilities of UAVs, can precisely characterize the boundaries of small water bodies such as narrow ditches and small ponds, achieving accurate identification and location of small-scale black and odorous water bodies.

[0032] 4. The monitoring and intelligent identification method and system for small and micro black and odorous water bodies provided by the present invention utilizes multiple key indicators such as transparency, ammonia nitrogen, and dissolved oxygen for collaborative judgment. Unlike the existing technology that mostly uses a single indicator or empirical index to classify water bodies, this method can automatically complete the classification and identification of normal water bodies as mildly or severely black and odorous, ensuring the consistency between monitoring results of different batches and improving operability in actual business scenarios.

[0033] 5. The monitoring and intelligent identification method and system for small and micro black and odorous water bodies provided by this invention constructs a complete "space-air-ground" integrated technology link, realizing a closed loop of the entire process from raw data acquisition, stable feature screening, high-resolution spatial reconstruction to business-level identification. It closely integrates remote sensing modeling and analysis with the business needs of water environment supervision, effectively improving the engineering feasibility of the method and enhancing the practical application value of the technology in the routine supervision of small and micro black and odorous water bodies. Attached Figure Description

[0034] Figure 1 This is a technical roadmap of the present invention. Detailed Implementation

[0035] The present invention will now be described in detail with reference to the accompanying drawings.

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0037] A monitoring and intelligent identification method for small and micro black and odorous water bodies, such as Figure 1 As shown, it includes the following steps: S1, Satellite-to-Ground Data Matching Phase: Radiometric calibration is performed on the hyperspectral satellite data to eliminate sensor system errors. The FLAASH model is then used to perform atmospheric correction on the radiometrically calibrated data to obtain the true water surface remote sensing reflectance after atmospheric interference is eliminated, resulting in the processed satellite layer. Prior to this step, it is necessary to obtain hyperspectral satellite remote sensing reflectance data within the monitoring area, taking the overpassing OHS hyperspectral satellite data as an example (containing 32 continuous hyperspectral bands with a spatial resolution of 10 meters). A rough water mask was extracted from the satellite layer using the normalized differential water index. For the pixels at the mask boundary, a linear spectral unmixing model was used to perform abundance inversion with water, vegetation and soil as endmembers. Only pixels with pure water abundance greater than the target value were retained to complete pixel purification and obtain the spectral reflectance curve of the satellite pixels. The purified satellite pixel spectral reflectance curves are spatially aligned with ground-synchronized measured data to construct a primary training sample library for satellite-ground integration. Prior to this step, ground-synchronized (time difference within ±2 days) measured data (including physicochemical indicators such as transparency SD, ammonia nitrogen NH3-N, and dissolved oxygen DO) needs to be collected. S2, Optimal Feature Mining Stage: Using purified hyperspectral satellite remote sensing reflectance and derived band indices as input variables and ground-measured water quality indicators as dependent variables, an integrated learning primary inversion model based on LightGBM was constructed to fit the nonlinear mapping pathway between high-dimensional spectral space and water quality indicators. Perform M rounds of Bootstrap self-sampling on the historical sample dataset, which contains T sample time phases; wherein, the number of sample time phases in the historical sample dataset is ≥30, and M≥100; In each round of sampling, the SHAP game theory framework is invoked to perform inverse analysis of the model, calculating the Shapley marginal contribution value of each spectral feature. The calculation expression is as follows: In the formula, Let F be the marginal contribution value of spectral feature i, F be the full-spectrum feature set, and S be the feature subset excluding feature i. This is the model's predicted output on the feature subset S; The absolute values ​​of sapride for each spectral characteristic under M rounds of sampling. The variance is calculated, and only the high-frequency intersection features with variances less than a set threshold and ranking in the top N of the multi-temporal sample sets are retained as the optimal features. These high-frequency intersection features are typically combinations of 3–5 core bands, and the optimal feature A in this scheme is the targeted core feature. This invention specifically defines the optimal features as not, in physical essence, equivalent to the black and odorous water body itself, but rather as a set of spectral representations that have statistically sensitive responses to target water quality deterioration indicators (ammonia nitrogen, dissolved oxygen, etc.). Furthermore, the optimal features used for subsequent spectral gradient morphology consistency constraints are limited to single-band reflectance features with a clear center wavelength correspondence, or spectral features composed of neighboring bands with a clear center wavelength correspondence and capable of being reconstructed to the reflectance of the corresponding band. This feature compression significantly replaces the traditional "blind full-spectrum simulation," converging the mapping solution space of the subsequent super-resolution reconstruction model.

[0038] S3, AI spectral super-resolution reconstruction stage: Construct a convolutional neural network model based on an encoder-decoder architecture; The drone data and the optimal features output by S2 are concatenated at the front end of the encoder of the convolutional neural network model to achieve multimodal feature concatenation. Before this step, drone data needs to be acquired, but it is not limited to which step it is acquired. The drone data is a centimeter-level high spatial resolution RGB three-channel orthophoto. In the backpropagation stage of training the convolutional neural network model, a loss function embedding physical constraints of spectral gradients is designed to constrain the reconstruction process. The loss function is: In the formula, This is the mean square error loss term. For spatial continuity constraints, For the morphological consistency constraint term of the spectroscopic gradient, and Weighting adjustment coefficient; A method and system for monitoring and intelligent identification of water bodies; In the formula, A is the optimal feature; i represents the row index of the image pixel; j represents the column index of the image pixel; ∑i,j represents the summation of the spectral gradient differences of each pixel position in the reconstructed region; ∥⋅∥ represents the norm used to measure the difference between the UAV reconstructed spectral gradient and the satellite prior spectral gradient. The spectral gradient morphological consistency constraint term The expression is: In the formula, A λa and A λbrespectively represent the feature values corresponding to adjacent or neighboring optimal spectral features with wavelengths of λa and λb; λa and λb respectively represent the central wavelengths of the wavebands corresponding to the two optimal spectral features; the subscript UAV represents the optimal feature with high spatial resolution obtained by AI spectral super-resolution reconstruction from UAV images; the subscript Sat_Prior represents the corresponding optimal spectral feature prior determined from hyperspectral satellite data; i represents the row index of image pixels; j represents the column index of image pixels; ∑i,j represents the summation of the spectroscopic gradient differences at each pixel position within the reconstruction area; ∥⋅∥ represents the norm used to measure the difference between the UAV reconstructed spectroscopic gradient and the satellite prior spectroscopic gradient.

[0039] The convolutional neural network model is subjected to training iterations to output an optimal feature space distribution map with high spatial resolution; in this embodiment, 2000–5000 training iterations are performed, and the initial learning rate is set to 10^-4, S4. Multi-indicator business classification and identification stage: The local centimeter-level optimal feature distribution map generated by AI spectral super-resolution reconstruction is used as input to map and output three spatial thematic maps of water quality indicators at centimeter-level resolution, namely transparency, ammonia nitrogen and dissolved oxygen; the three spatial thematic maps of water quality indicators are mapped to the fixed threshold discrimination matrix of standard administrative business specifications; In the fixed threshold discrimination matrix, the three indicators of transparency, ammonia nitrogen and dissolved oxygen each correspond to a preset normal water body judgment threshold, light black and odorous water judgment threshold, and heavy black and odorous water judgment threshold; specifically, the matrix strictly hard-codes the core control parameters of standards such as the "Guidelines for the Renovation of Rural Black and Odorous Water Bodies" promulgated by the Ministry of Ecology and Environment of China, and the standards are as follows: The judgment interval for transparency SD is: normal water body SD≥25cm, light black and odorous 10cm≤SD<25cm, heavy black and odorous SD<10cm; The judgment interval for ammonia nitrogen NH3-N is: normal water body NH3-N≤2.0mg / L, light black and odorous 2.0mg / L<NH3-N≤8.0mg / L, heavy black and odorous NH3-N>8.0mg / L; The judgment interval for dissolved oxygen DO is: normal water body DO≥2.0mg / L, light black and odorous 0.2mg / L≤DO<2.0mg / L, heavy black and odorous DO<0.2mg / L; The fixed threshold discrimination matrix performs a pixel-by-pixel single-factor one-vote veto cascade scan. If any water quality indicator of any pixel triggers its corresponding pollution judgment threshold, the pixel is judged as the corresponding grade; if none of all the indicators trigger the pollution judgment threshold, it is judged as a normal water body.

[0040] An integrated sky-ground monitoring and intelligent identification system, comprising: The data preprocessing module standardizes the hyperspectral satellite data and outputs the processed satellite layers. The data matching module aligns the processed satellite layer with the measured point data using spatial grid coordinates and performs pixel purification to construct a primary training sample library integrating satellite and ground. The primary inversion model for water quality parameters at the space-to-ground scale uses data from the space-to-ground integrated primary training sample library as input to establish the mapping relationship between spectral features and water quality indicators, and outputs the mapping model between spectral features and water quality indicators. The SHAP evaluation module performs dimensionality reduction and screening on the mapping relationship output by the primary inversion model of water quality parameters at the terrestrial and space scales by quantifying the marginal contribution of each spectral feature, and outputs the optimal features that are stable and reproducible. The AI ​​spectral super-resolution reconstruction module takes UAV data and optimal features as bidirectional input variables, and outputs a local optimal feature distribution map through super-resolution reconstruction processing with embedded spectral gradient physical constraints. The UAV-scale water quality inversion model uses the local optimal feature distribution map as input to invert a high spatial resolution water quality parameter distribution map. The multi-indicator collaborative water body business classification mapping module maps the water quality parameter distribution map to a fixed threshold discrimination matrix of standard administrative business specifications; The fixed threshold discrimination matrix module performs multi-level logic gate judgments on the mapped water quality parameter distribution map, and outputs different business classification results respectively.

[0041] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. The invention extends to any new features or combinations disclosed in this specification, and any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the above exemplary embodiments.

Claims

1. A method for monitoring and intelligent identification of small-scale black and odorous water bodies, characterized in that, Includes the following steps: S1. Satellite-Ground Data Matching Stage: Data preprocessing is performed on hyperspectral satellite data to obtain processed satellite layers. Data matching is performed between the processed satellite layers and the measured point data to complete spatial grid coordinate alignment and pixel purification. S2, Optimal Feature Mining Stage: After completing data matching, the data is input into the primary inversion model of water quality parameters at the space-to-ground scale to establish the mapping relationship between spectral features and water quality indicators; and the SHAP evaluation with stability constraint mechanism is introduced to perform dimensionality reduction screening by quantifying the marginal contribution of each spectral feature, and output the stable and reproducible optimal features. S3, AI Spectral Super-Reconstruction Stage: The UAV data and the optimal features output from S2 are used as bidirectional input variables to jointly perform AI spectral super-reconstruction and output a local optimal feature distribution map. S4, Multi-indicator Business Classification and Identification Stage: The local optimal feature distribution map output from S3 is used as input to inject and construct the UAV-scale water quality inversion model, and a high spatial resolution water quality parameter distribution map is inverted. The water quality parameter distribution map is then used for multi-index collaborative water body business classification mapping. Classification is determined by a fixed threshold discrimination matrix, and independent business classification results are output for each classification.

2. The method as described in claim 1, characterized in that, The data preprocessing in S1 includes: Radiometric calibration was performed on the hyperspectral satellite data to eliminate sensor system errors; the FLAASH model was used to perform atmospheric correction on the radiometrically calibrated data to obtain the true remote sensing reflectance of the water surface after eliminating atmospheric interference.

3. The method as described in claim 1, characterized in that, The data matching in S1 includes: A coarse water mask was extracted from the satellite layer using the normalized differential water index. For pixels at the mask boundary, a linear spectral unmixing model was used to perform abundance inversion with water, vegetation, and soil as endmembers. Only pixels with pure water abundance greater than the target value were retained to complete pixel purification and obtain the satellite pixel spectral reflectance curve. The purified satellite pixel spectral reflectance curve was then spatially aligned with the ground-synchronized measured point data to construct a satellite-ground integrated primary training sample library.

4. The method as described in claim 1, characterized in that, The method for constructing the primary inversion model of water quality parameters at the terrestrial and space scales in S2 includes: Using purified hyperspectral satellite remote sensing reflectance in each band as input independent variables and ground-measured water quality indicators as dependent variables, an ensemble learning primary inversion model based on LightGBM is constructed to fit the nonlinear mapping pathway between high-dimensional spectral space and water quality indicators.

5. The method as described in claim 1, characterized in that, The SHAP evaluation that introduces a stability constraint mechanism in S2 specifically includes: A1. Perform M rounds of Bootstrap self-sampling on the historical sample dataset, which contains T sample time phases; A2. In each round of sampling, the SHAP game theory framework is used to perform inverse analysis of the model, calculating the Shapley marginal contribution value of each spectral feature. The calculation expression is as follows: In the formula, Let F be the marginal contribution value of spectral feature i, F be the full-spectrum feature set, and S be the feature subset excluding feature i. This is the model's predicted output on the feature subset S; A3. Statistical analysis of the absolute values ​​of sapride for each spectral characteristic under M rounds of sampling. The variance is used to retain only the high-frequency intersection features whose variance is less than a set threshold and which rank in the top N of the multi-temporal sample sets, and these features are established as the optimal features.

6. The method as described in claim 1, characterized in that, The AI ​​spectral super-resolution reconstruction in S3 specifically includes: B1. Construct a convolutional neural network model based on an encoder-decoder architecture; B2. The UAV data and the optimal features output by S2 are concatenated at the front end of the encoder of the convolutional neural network model to achieve multimodal feature concatenation. B3. During the backpropagation stage of the convolutional neural network model training, a loss function embedding physical constraints of spectral gradients is designed to constrain the reconstruction process. The loss function is: In the formula, This is the mean square error loss term. For spatial continuity constraints, For the morphological consistency constraint term of the spectroscopic gradient, and This is the weighting adjustment coefficient; B4. After training and iterating the convolutional neural network model, output the optimal feature space distribution map with high spatial resolution.

7. The method as described in claim 6, characterized in that, The spatial continuity constraint term The expression is: In the formula, A is the optimal feature; i represents the row index of the image pixel; j represents the column index of the image pixel; ∑i,j represents the summation of the spectral gradient differences of each pixel position in the reconstructed region; ∥⋅∥ represents the norm used to measure the difference between the UAV reconstructed spectral gradient and the satellite prior spectral gradient. The spectral gradient morphological consistency constraint term The expression is: In the formula, A λa and A λb These represent the eigenvalues ​​corresponding to the adjacent or neighboring optimal spectral features with wavelengths λa and λb, respectively; λa and λb represent the center wavelengths of the bands corresponding to the two optimal spectral features, respectively; the subscript UAV indicates the high spatial resolution optimal feature obtained by AI spectral super-resolution reconstruction of UAV imagery; the subscript Sat_Prior indicates the prior of the corresponding optimal spectral feature determined by hyperspectral satellite data; i represents the row index of the image pixel; j represents the column index of the image pixel; ∑i,j represents the summation of the spectral gradient differences at each pixel location within the reconstructed region; ∥⋅∥ represents the norm used to measure the difference between the UAV reconstructed spectral gradient and the satellite prior spectral gradient.

8. The method as described in claim 1, characterized in that, The method for constructing the UAV-scale water quality inversion model in S4 includes: The local centimeter-level optimal feature distribution map generated by AI spectral super-resolution reconstruction is used as input to map and output spatial thematic maps of three water quality indicators, namely transparency, ammonia nitrogen, and dissolved oxygen, at centimeter resolution. The spatial thematic maps of the three water quality indicators are then mapped to a fixed threshold discrimination matrix in the standard administrative business specifications.

9. The method as described in claim 8, characterized in that, The determination rule for the fixed threshold discrimination matrix is ​​as follows: In the fixed threshold discrimination matrix, the three indicators of transparency, ammonia nitrogen and dissolved oxygen each have a preset threshold for normal water body judgment, a threshold for mild black and odorous water body judgment and a threshold for severe black and odorous water body judgment. The fixed threshold discrimination matrix performs a pixel-by-pixel single-factor veto cascade scan. If any water quality indicator of any pixel triggers its corresponding pollution judgment threshold, the pixel is judged as the corresponding level. If none of the indicators trigger the pollution judgment threshold, the pixel is judged as normal water.

10. A sky-ground integrated monitoring and intelligent identification system, characterized in that, include: The data preprocessing module standardizes the hyperspectral satellite data and outputs the processed satellite layers. The data matching module aligns the processed satellite layer with the measured point data using spatial grid coordinates and performs pixel purification to construct a primary training sample library integrating satellite and ground. The primary inversion model for water quality parameters at the space-to-ground scale uses data from the space-to-ground integrated primary training sample library as input to establish the mapping relationship between spectral features and water quality indicators, and outputs the mapping model between spectral features and water quality indicators. The SHAP evaluation module performs dimensionality reduction and screening on the mapping relationship output by the primary inversion model of water quality parameters at the terrestrial and space scales by quantifying the marginal contribution of each spectral feature, and outputs the optimal features that are stable and reproducible. The AI ​​spectral super-resolution reconstruction module takes UAV data and optimal features as bidirectional input variables, and outputs a local optimal feature distribution map through super-resolution reconstruction processing with embedded spectral gradient physical constraints. The UAV-scale water quality inversion model uses the local optimal feature distribution map as input to invert a high spatial resolution water quality parameter distribution map. The multi-indicator collaborative water body business classification mapping module maps the water quality parameter distribution map to a fixed threshold discrimination matrix of standard administrative business specifications; The fixed threshold discrimination matrix module performs multi-level logic gate judgments on the mapped water quality parameter distribution map, and outputs different business classification results respectively.

Images