Deep learning driven remote sensing identification and analysis system and method for lake water turbidity

Abstract

The application discloses a kind of deep learning driving lake water turbidity remote sensing identification analysis system and method, the system includes: multi-source heterogeneous data preprocessing unit, water body optical type partition unit, partition adaptation deep learning turbidity inversion unit and turbidity space-time correlation analysis unit.The method adopts cross-validation residual error driven local iterative clustering collaborative inversion method, through partition attribution residual error screening, residual connected domain identification, construction adjustment candidate partition and inversion precision-complexity linkage evaluation, realize lake water body optical partition boundary and turbidity inversion model linkage iterative optimization.The application can fully exploit lake area turbidity inversion difference information, accurately match model application range, reduce partition transition area inversion error, effectively improve the scene adaptation level and global inversion accuracy of turbidity remote sensing identification under complex lake environment.

Description

Technical Field

[0001] This invention relates to the field of remote sensing monitoring technology for inland lake water environment, and in particular to a deep learning-driven remote sensing identification and analysis system and method for lake water turbidity. Background Technology

[0002] With the development of remote sensing technology, artificial intelligence technology, and multi-source data fusion technology, monitoring of lake water environmental parameters based on satellite remote sensing imagery is gradually being applied in the field of water environment monitoring. By acquiring the spectral information of water bodies in remote sensing images and combining it with ground-measured data, the correlation between remote sensing image features and water parameters can be established, enabling spatial monitoring of the lake water environment status and providing data support for lake ecological environment management and water resource protection.

[0003] In the prior art, Chinese patent document CN202110601951.4 discloses a method and system for calculating the transparency of nearshore and inland water bodies. This method collects spectral data of nearshore and inland water bodies, performs cluster analysis on the spectral data, classifies the water bodies into different types based on the analysis results, selects a corresponding preset transparency inversion model for each water body type, and uses the preset transparency inversion model to calculate the water transparency from the spectral data. Chinese patent document CN202110633489.6 discloses a remote sensing water quality inversion method combining spatiotemporal fusion and deep learning. This method generates high spatiotemporal resolution water quality inversion results by combining a spatiotemporal fusion model with a fully connected neural network and a convolutional neural network, and uses a pixel decomposition method to correct the water quality inversion results, thereby improving the spatial and temporal resolution of the remote sensing image water quality inversion results.

[0004] Although the aforementioned technical solutions have advanced the development of remote sensing monitoring technology for water parameters from aspects such as water body optical classification and multi-source remote sensing deep learning inversion, two technical problems still urgently need to be solved in the actual application scenarios of complex lake water bodies: First, the adaptability of lake water optical zoning and inversion models is insufficient, and static zoning methods are difficult to dynamically match the optimal boundary of inversion accuracy. Chinese patent document CN202110601951.4 adopts a static two-stage architecture of one-time boundary determination by spectral clustering and independent modeling of zoning. The zoning boundary is determined only based on the initial spectral features, and there is no interactive feedback mechanism with the subsequent inversion process; once the zoning result is generated, it remains fixed and cannot dynamically adjust the boundary range according to the inversion execution effect. When the optical spatial heterogeneity of lake water is strong, the fixed zoning boundary is difficult to accurately match the optimal applicable range of the inversion model, which easily leads to high inversion error in the transition area of ​​the zoning, and cannot guarantee the consistency and stability of the inversion accuracy of the entire lake. Second, the spatial difference information in the inversion process does not form an effective feedback link, and regional inversion errors cannot be used to optimize the zoning and modeling strategies. Chinese patent document CN202110633489.6 adopts multi-source data spatiotemporal fusion and neural network model architecture optimization, but only uses the inversion residual as a post-event evaluation index of model accuracy, without establishing a reverse feedback mechanism for spatial differences in inversion results; it cannot use the spatial distribution characteristics of inversion residuals to guide the adjustment of zonal boundaries and model adaptation optimization, and it is difficult to further improve the accuracy of water quality parameter inversion and scenario adaptation capability in complex water environments by mining regional inversion difference information. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a deep learning-driven remote sensing identification and analysis system and method for lake water turbidity, in order to solve the problems of insufficient adaptability between lake water optical zoning and inversion models, difficulty in dynamically matching the boundary of the region with the best turbidity inversion accuracy using static zoning methods, lack of effective feedback links for spatial difference information in the inversion process, and inability to reverse-optimize zoning and modeling strategies due to regional turbidity inversion errors.

[0006] Technical solution: A deep learning-driven remote sensing identification and analysis system for lake water turbidity, comprising:

[0007] The multi-source heterogeneous data preprocessing unit is used to acquire multi-source satellite remote sensing images, measured turbidity data of lake water bodies, and meteorological and hydrological auxiliary data. It performs radiometric correction, atmospheric correction, water body area extraction, and spatiotemporal registration of remote sensing images and measured data. It also performs standardized preprocessing on the acquired raw data to obtain water body spectral datasets, measured turbidity sample datasets, and meteorological and hydrological auxiliary datasets.

[0008] The water body optical type partitioning unit receives water body spectral datasets and uses spectral feature extraction and cluster analysis methods to perform optical type partitioning on the entire lake area, obtaining initial water body optical parent partitioning results. It includes: a water body spectral feature extraction module, which receives the water body spectral dataset, extracts the ratio of single-band reflectance to combined bands, and generates a water body spectral feature set; a water body optical clustering module, which receives the water body spectral feature set, uses cluster analysis methods to perform optical type clustering on the entire lake area, and obtains preliminary water body optical partitioning results; and a parent partitioning result adjustment module, which receives the preliminary water body optical partitioning results, performs spatial connectivity regularization on the partition boundaries, and removes patches with areas smaller than a preset threshold, obtaining the initial water body optical parent partitioning results.

[0009] The partition adaptation deep learning turbidity inversion unit is used to receive the initial water body optical parent partition results, water body spectral dataset, and measured turbidity sample dataset. It adopts a local iterative clustering collaborative inversion method driven by cross-validation sample residuals to perform linkage iterative optimization processing on the partition boundaries and turbidity inversion model within each optical parent partition, so as to obtain the turbidity inversion adaptation sub-partition results and the spatial distribution results of turbidity in the entire lake area.

[0010] The turbidity spatiotemporal correlation analysis unit is used to receive the spatial distribution results of turbidity in the entire lake area, the results of turbidity inversion and adaptation sub-regions, and meteorological and hydrological auxiliary datasets. It uses knowledge graph and correlation factor mining methods to analyze and process the spatiotemporal evolution law and driving factors of lake turbidity, and obtain the turbidity change trend and driving factor analysis results.

[0011] Furthermore, the partition-adaptive deep learning turbidity inversion unit includes a cross-validation sample residual generation module, a partition attribution sample residual screening module, a local clustering boundary adjustment module, and a partition inversion linkage iteration module;

[0012] The cross-validation sample residual generation module receives the initial water body optical parent partition results, water body spectral dataset, and measured turbidity sample dataset. It organizes the measured turbidity sample dataset and the corresponding water body spectral data according to the initial water body optical parent partition and generates a cross-validation inversion residual sample point set.

[0013] The partition attribution sample residual screening module receives the cross-validation inversion residual sample point set, removes error components that are not caused by partitioning, and obtains candidate partition attribution residual samples.

[0014] The local clustering boundary adjustment module receives candidate partition attribution residual samples and initial water body optical parent partition results, and generates local partition adjustment candidates within the corresponding water body optical parent partition;

[0015] The partition inversion linkage iteration module receives local partition adjustment candidates, determines the validity of partition adjustment and controls the iteration process. After the iteration terminates, it outputs the turbidity inversion results of the adapted sub-partitions and the spatial distribution results of turbidity inversion of the entire lake area.

[0016] Furthermore, the turbidity spatiotemporal correlation analysis unit includes a partitioned time series data integration module, a driving factor knowledge graph construction module, and a spatiotemporal evolution and driving analysis module;

[0017] The partitioned time-series data integration module receives the spatial distribution results of turbidity inversion of the entire lake area, the turbidity inversion results of the adapted sub-partitions, and the meteorological and hydrological auxiliary datasets. According to the boundaries of the adapted sub-partitions, it integrates the time-series turbidity data of each sub-partition with the corresponding meteorological and hydrological data to generate the partitioned time-series dataset.

[0018] The knowledge graph construction module for driving factors receives partitioned time-series datasets, establishes the association between turbidity and meteorological and hydrological element entities, and constructs a knowledge graph of lake turbidity driving factors.

[0019] The spatiotemporal evolution and driving analysis module receives a knowledge graph of lake turbidity driving factors, mines driving factors that meet preset contribution thresholds in the turbidity changes of each sub-region, analyzes the spatiotemporal evolution law of turbidity, and obtains the turbidity change trend and driving factor analysis results.

[0020] A deep learning-driven remote sensing identification and analysis method for lake water turbidity, based on any of the analysis systems described above, realizes the analysis of the trend of turbidity changes in lake water, including the following steps:

[0021] S1: The multi-source raw data acquisition module acquires multi-source satellite remote sensing images, measured turbidity data of lake water, and meteorological and hydrological auxiliary data, and preprocesses the acquired raw data to obtain water spectral dataset, measured turbidity sample dataset, and meteorological and hydrological auxiliary dataset.

[0022] S2, the reflectance, band ratio, and spectral peak-valley characteristics of turbidity-sensitive characteristic bands in the water body optical type partitioning unit are extracted from the water body spectral data, and the reflectance, band ratio, and spectral peak-valley characteristics are combined to form a water body spectral feature set; then, cluster analysis is performed on the water body spectral feature set, and each water body pixel in the entire lake area is assigned a category label to represent the optical type to which the pixel belongs; the set of pixels with the same category label and spatially adjacent constitutes a preliminary water body optical partition, and all water body optical partitions together constitute the preliminary water body optical partitioning result;

[0023] Next, spatial connectivity regularization is performed on the partition boundaries, and patches with an area smaller than a preset threshold are merged to obtain the initial water body optical parent partition results.

[0024] S3, using a cross-validation residual-driven local iterative clustering collaborative inversion method, the partition boundary of the water body optical parent partition and the turbidity inversion model are linked and iteratively optimized to obtain the turbidity inversion results of the adapted sub-partitions and the spatial distribution results of turbidity in the entire lake area.

[0025] S4, the turbidity spatiotemporal correlation analysis unit, adopts knowledge graph construction and correlation factor mining methods. Based on the turbidity inversion results of the adaptive sub-region, the spatial distribution results of turbidity in the whole lake area, and the meteorological and hydrological auxiliary dataset, it analyzes and processes the spatiotemporal evolution law and driving factors of lake turbidity, and obtains the turbidity change trend and driving factor analysis results.

[0026] Furthermore, in step S2, the step of performing spatial connectivity regularization on the partition boundaries includes: the parent partition result regularization module performs connectivity analysis on the pixel set of each optical type using the eight-neighbor connectivity criterion to identify each independent connected component; for connected components with larger areas and more regular boundaries, they are directly retained as valid partitions; for jagged or irregular protrusions on the boundaries caused by cluster boundary jitter, morphological smoothing or boundary refinement based on spatial neighborhood relationships is used to make the partition boundaries more regular and smooth.

[0027] The steps for merging patches with an area smaller than a preset threshold include: the parent partition result normalization module calculates the area of ​​each connected component and re-merges patches with an area smaller than the preset area threshold into the optical type partition with the largest spatially adjacent area; wherein, the preset area threshold is determined based on the spatial resolution of the remote sensing image and the size of the lake area.

[0028] Furthermore, in step S3, the steps to obtain the turbidity inversion results of the adaptor sub-region and the spatial distribution results of turbidity across the entire lake include:

[0029] S31, the cross-validation sample residual generation module uses a cross-validation residual-driven local iterative clustering collaborative inversion method to generate a set of cross-validation inversion residual sample points, including:

[0030] S311, group the measured turbidity sample dataset and the corresponding water body spectral data according to the initial water body optical parent partition, forming the sample subset corresponding to each water body optical parent partition;

[0031] S312, For each sample subset, cross-validation is used to divide the training group and the validation group, and the turbidity inversion model of the corresponding partition is constructed based on the training group data;

[0032] S313, calculate the predicted turbidity value of the corresponding validation group sample through the turbidity inversion model of each partition, and obtain the inversion residual of each sample by combining the corresponding measured turbidity value, forming a cross-validation inversion residual point set with spatial location and spectral attributes;

[0033] S32, the partition attribution sample residual screening module filters the cross-validation inversion residual sample point set to obtain candidate partition attribution residual samples; including:

[0034] S321, perform error standardization on the set of cross-validation inversion residuals in each optical parent partition to unify the dimensional scale of the residuals;

[0035] S322, count the common offset residual components of the residuals in each optical parent partition, remove the corresponding common offset residual components from each sample residual, and obtain the residual point set after system bias correction;

[0036] S323, Spatial filtering is performed on the corrected residual point set to remove isolated residual samples from the corrected residual point set, and candidate partition attribution residual samples are obtained.

[0037] S33, the local clustering boundary adjustment module processes the candidate partition attribution residual samples and the initial water body optical parent partition results, generating local partition adjustment candidates within the corresponding optical parent partition; including:

[0038] S331, For the region corresponding to the residual connected component, extend the sub-partition boundary within the scope of the single optical parent partition to form a partition splitting candidate; and treat the partition splitting candidate as a new sub-partition.

[0039] S332, generate partition merging candidates for spatially adjacent sub-partitions within the same water body optical parent partition;

[0040] S333, perform spectral feature separability verification on all partition splitting and partition merging candidates, and use the candidates that pass the verification as the final local partition adjustment candidates;

[0041] S34, the partition inversion linkage iteration module determines the effectiveness of local partition adjustment candidates, controls the iteration process based on the adjustment effectiveness determination result, and updates the sub-partitions affected by the partition adjustment and the turbidity inversion model; after the iteration terminates, it outputs the turbidity inversion results of the adapted sub-partitions and the spatial distribution results of turbidity in the whole lake area.

[0042] Furthermore, the turbidity inversion model adopts a multilayer perceptron structure. The number of nodes in the input layer is set to 6, corresponding to the reflectance of the blue light band, the reflectance of the green light band, the reflectance of the red light band, the reflectance of the near-infrared band, the ratio of the green light band to the blue light band, and the ratio of the near-infrared band to the red light band, respectively. The number of hidden layers is set to 2, with 64 nodes in the first hidden layer and 32 nodes in the second hidden layer. The activation function for both layers is the ReLU function. The number of nodes in the output layer is set to 1, directly outputting the predicted turbidity value.

[0043] Furthermore, in step S34, the cross-validation inversion error and partition complexity before and after adjustment are calculated for the corresponding regions of the local partition adjustment candidates, and an adjustment evaluation index is generated by combining the accuracy gain and complexity cost; wherein, the adjustment evaluation index satisfies the following relationship:

[0044] ;

[0045] In the formula, In order to adjust the evaluation indicators, The amount by which the cross-validation inversion error is reduced before and after partitioning adjustment. The increment in partition complexity before and after partition adjustment. This is a preset complexity cost coefficient;

[0046] Next, based on the adjusted evaluation indicators, the results of the adjustment effectiveness determination are obtained: when When the local partition adjustment candidate is deemed a valid adjustment, it is adopted; when If this occurs, it indicates that the local partition adjustment candidate is an invalid adjustment and is rejected.

[0047] Finally, based on the results of the adjustment effectiveness judgment, the iteration process is controlled and only the sub-partitions and turbidity inversion models affected by the partition adjustment are updated. After the iteration terminates, the final partition scheme is output as the turbidity inversion adaptation sub-partition result. Using the turbidity inversion model of each adaptation sub-partition, the spectral data of all water pixels in the corresponding sub-partition are inverted pixel by pixel to obtain the turbidity prediction value of each pixel, and a turbidity spatial distribution map covering the entire lake area is generated as the turbidity spatial distribution result of the entire lake area.

[0048] Furthermore, in step S4, the specific steps for obtaining the turbidity change trend and driving factor analysis results include:

[0049] S41, the partitioned time-series data integration module integrates the time-series turbidity data and corresponding meteorological and hydrological data of each sub-partition according to the boundaries of the adapted sub-partitions, generating a partitioned time-series dataset; including:

[0050] The final optimized spatial boundaries of each sub-partition are obtained from the turbidity inversion adaptation sub-partition results.

[0051] Based on the spatial boundaries of each sub-region, the spatial distribution data of turbidity in each time phase is statistically analyzed by sub-region; the average value of the turbidity of all water pixels in each sub-region is taken as the turbidity value of that sub-region in that time phase; after processing all time phases in sequence, the time-series turbidity sequence of each sub-region is obtained.

[0052] The meteorological and hydrological data from multiple automatic observation stations within each sub-region are averaged and correlated with the turbidity of the corresponding sub-region to obtain the meteorological and hydrological element value sequence of each sub-region.

[0053] After completing the data association, the time-series turbidity sequence of each sub-partition is aligned and integrated with the meteorological and hydrological element sequence of the corresponding time period according to a unified time axis;

[0054] After integration, each sub-partition forms a partition time-series data record, which includes the sub-partition number, timestamp sequence, turbidity time-series sequence, and time-series sequences of meteorological and hydrological elements; the partition time-series data records of all sub-partitions are summarized to form a partition time-series dataset.

[0055] S42, Based on the partitioned time-series dataset, extract entities related to lake turbidity changes and the relationships between these entities, organize and store them according to a graph structure, and construct a knowledge graph of lake turbidity driving factors; details are as follows:

[0056] Three types of entities are identified: sub-region entities, where each turbidity inversion adaptation sub-region is treated as an independent entity, with attributes including sub-region number, spatial extent, area, and optical type; turbidity state entities, where the turbidity values ​​of each sub-region at each time phase are discretized into different turbidity state levels according to a preset classification standard, and each turbidity state entity includes a level label, turbidity value range, and corresponding timestamp; and meteorological and hydrological event entities, which identify specific event types from the time series of meteorological and hydrological elements.

[0057] Two types of relationships are constructed: spatial affiliation relationship, which connects sub-partition entities and turbidity state entities, indicating that a certain turbidity state occurs within a certain sub-partition; and temporal driving relationship, which connects meteorological and hydrological event entities and turbidity state entities, indicating that a certain meteorological and hydrological event occurs before a certain turbidity state in time.

[0058] Generate a knowledge graph of lake turbidity driving factors: nodes represent entities, edges represent relationships, and both nodes and edges are attached with attribute information;

[0059] S43, Based on the knowledge graph of lake turbidity driving factors, mine the driving factors that satisfy the preset contribution threshold for turbidity changes in each sub-region, analyze the spatiotemporal evolution law of turbidity, and obtain the turbidity change trend and driving factor analysis results; including:

[0060] Analysis of the spatiotemporal evolution of turbidity: The spatiotemporal evolution and driving analysis module extracts the turbidity status information of each sub-region at each time phase from the knowledge graph of lake turbidity driving factors, and analyzes the turbidity change trend of each sub-region along the time axis; at the same time, by comparing the turbidity change trends of different sub-regions in the same period, the spatial differentiation characteristics of turbidity change within the lake are identified.

[0061] Driving Factor Mining: The Spatiotemporal Evolution and Driving Analysis module utilizes the temporal driving relationship between meteorological and hydrological events and turbidity status stored in the knowledge graph of lake turbidity driving factors to quantitatively assess the contribution of each meteorological and hydrological element to the turbidity changes in each sub-region; and determines the main driving factors of turbidity changes in each sub-region based on the contribution level.

[0062] For each sub-partition, the turbidity change trend and the main driving factors are listed, and the results of the turbidity change trend and driving factor analysis are summarized.

[0063] Furthermore, the results of the analysis of turbidity change trends and driving factors are presented in the form of tables or thematic maps: the tables list the number, area, direction and rate of turbidity change trends, main driving factors and their contribution to each sub-region; the thematic maps use different colors or symbols on the lake zoning map to represent the type of turbidity change trend and the category of main driving factors for each sub-region.

[0064] Compared with the prior art, the significant advantages of this invention are as follows:

[0065] 1. This invention adopts a cross-validation residual-driven local iterative clustering collaborative inversion method. By screening the attribution residuals of the whole lake area and dynamically adjusting the local boundaries, it realizes the linkage iterative optimization of the lake area boundary and the turbidity inversion model. It solves the limitation of static partitioning and one-time boundary determination, can accurately match the optimal applicable range of the model, reduce the inversion error in the transition area of ​​the partition, and ensure the stability of the turbidity inversion accuracy of the whole lake area in complex lakes.

[0066] 2. The turbidity inversion residual spatial distribution reverse feedback optimization link constructed in this invention adopts the method of residual connected component identification and accuracy-complexity linkage evaluation mechanism to control the iteration process, explore the optimization value of regional turbidity inversion difference information, open up the transmission path of reverse optimization of partitioning and modeling by inversion error, and improve the adaptability and overall accuracy of turbidity remote sensing identification in complex lake water environment.

[0067] 3. The deep learning-driven remote sensing identification and analysis system for lake water turbidity of the present invention takes multi-source satellite remote sensing images, ground-measured data of lake water turbidity, and meteorological and hydrological auxiliary data as inputs, and takes the spatial distribution of lake turbidity inversion, turbidity change trend, and driving factor analysis results related to meteorological and hydrological data as outputs. Through multi-source heterogeneous data preprocessing unit, water body optical type partitioning unit, partition-adaptive deep learning turbidity inversion unit, and turbidity spatiotemporal correlation analysis unit, it realizes a fully automatic, rapid, and integrated function from real-time reception of satellite images and meteorological and hydrological data, through data standardization processing and water body turbidity inversion, to output of turbidity time series change and driving factor analysis, which can provide convenient services for lake basin water environment management. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the overall system framework structure of the analysis system of the present invention;

[0069] Figure 2 This is a flowchart of the analysis method of the present invention;

[0070] Figure 3 This is a schematic diagram of the lake dynamic optical zoning and turbidity model optimization inversion of the present invention. Detailed Implementation

[0071] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0072] like Figure 1 As shown, this embodiment provides a deep learning-driven remote sensing identification and analysis system for lake water turbidity, applied to the dynamic remote sensing monitoring of turbid water bodies such as large inland lakes and reservoirs. This analysis system relies on a conventional computer system, taking multi-source satellite remote sensing imagery, ground-based measured water turbidity data, and meteorological and hydrological auxiliary data as inputs. After data standardization, initial optical partitions of the lake water body are formed based on the characteristic spectral clustering of water turbidity. Then, by utilizing the spatial clustering characteristics of the cross-validation inversion residuals within each partition, model adaptation biases caused by partition boundaries are identified. These biases drive local adjustments to partition boundaries and corresponding updates to the turbidity inversion model. Through iterative processes, adapted sub-partitions that balance the optical separability of the water body with turbidity inversion accuracy are obtained. The turbidity of each sub-partition is inverted, and finally, the spatial distribution of turbidity across the entire lake, the turbidity variation trends of each partition in multiple time phases, and the driving factors affecting the turbidity changes in the lake water body are output.

[0073] In this embodiment, the analysis system is linked by a standardized dataset: the multi-source heterogeneous data preprocessing unit performs radiometric correction, water body extraction, and spatiotemporal registration on the raw satellite image data; the water body optical type partitioning unit generates an initial optical parent partition for lake water bodies from the preprocessed satellite images; the partition adaptation deep learning turbidity inversion unit performs iterative optimization by combining the water body spectrum of the satellite images, measured turbidity, and the parent partition results, outputting the adapted turbidity spatial distribution for lake sub-partitions and the entire region; and the turbidity spatiotemporal correlation analysis unit analyzes the turbidity change trend and driving factors by combining meteorological and hydrological data. Specifically:

[0074] The multi-source heterogeneous data preprocessing unit is used to acquire multi-source satellite remote sensing images of lakes, ground-measured turbidity data of water bodies, and meteorological and hydrological auxiliary data. Through radiometric correction of satellite images, water body region extraction, and spatiotemporal registration of measured turbidity data, it completes the standardized preprocessing of the acquired raw data, obtaining water body spectral datasets, measured turbidity sample datasets, and meteorological and hydrological auxiliary datasets. This unit includes a multi-source raw data acquisition module, a remote sensing image preprocessing module, and a multi-source data spatiotemporal registration module, wherein:

[0075] The multi-source raw data acquisition module automatically and in real time receives various domestic and international satellite remote sensing images, water turbidity data from automatic lake observation stations, and meteorological and hydrological auxiliary data through the API interface provided by the satellite website. It also inputs various satellite remote sensing images into the remote sensing image preprocessing module, and simultaneously inputs the measured lake water turbidity data and meteorological and hydrological data into the multi-source data spatiotemporal registration module.

[0076] The remote sensing image preprocessing module receives multi-source satellite remote sensing images of the lake. Since the data distribution agency has already performed geometric correction, it sequentially performs radiometric correction, atmospheric correction, and water area extraction of the images. Based on the complexity of the water in different areas of the lake, it sets sample points and the latitude and longitude coordinates of the lake's automatic observation station. It extracts the reflectance values ​​of all corresponding positions in all bands from the satellite images to obtain the initial water spectral samples, which are then input into the multi-source data spatiotemporal registration module.

[0077] The multi-source data spatiotemporal registration module receives initial water body spectral sample data, measured lake water turbidity data, and meteorological and hydrological auxiliary data. It completes the spatiotemporal matching and standardization of multi-source data through a unified spatial coordinate benchmark and matching time window, resulting in a spatiotemporally matched water body spectral dataset, measured turbidity sample dataset, and meteorological and hydrological auxiliary dataset.

[0078] The water body optical type partitioning unit includes a water body spectral feature extraction module, a water body optical clustering module, and a parent partitioning result regularization module;

[0079] The water body spectral feature extraction module receives the water body spectral dataset and obtains three types of spectral feature quantities that reflect the optical properties of the water body: single-band reflectance, ratio of each band, and spectral peak and valley characteristics, and generates a water body spectral feature set.

[0080] The water body optical clustering and classification module receives the water body spectral feature set as training samples and uses cluster analysis method to perform water body optical clustering on the remote sensing image of the entire lake area to obtain the preliminary water body optical zoning results of the lake.

[0081] The parent partitioning result normalization module receives the preliminary optical partitioning results of the lake, performs spatial connectivity normalization on the partition boundaries, and performs neighbor merging on patches with an area smaller than a preset threshold to obtain the initial optical parent partitioning results of the water body.

[0082] The partition-adaptive deep learning turbidity inversion unit includes a cross-validation sample generation module, a partition attribution sample residual screening module, a local clustering boundary adjustment module, and a partition inversion linkage iteration module.

[0083] The cross-validation sample generation module receives the initial optical parent partitioning results of the lake, the water spectral dataset, and the ground-measured turbidity sample dataset. Based on the initial optical parent partitioning of the lake, it divides the ground-measured turbidity sample dataset and the water spectral data within each partition and generates the cross-validation sample dataset for each partition.

[0084] The partition attribution sample residual screening module receives the cross-validation sample dataset, removes error components that are not caused by partitioning, and obtains candidate partition data samples.

[0085] The local clustering boundary adjustment module receives candidate partition data samples and the initial water body optical parent partition results, and generates local partition adjustment candidates within the corresponding optical parent partition;

[0086] The partition inversion linkage iteration module receives local partition adjustment candidates, determines the validity of partition adjustment and controls the iteration process. After the iteration terminates, it outputs the turbidity inversion adaptation sub-partition results and the spatial distribution results of turbidity in the entire lake area.

[0087] The turbidity spatiotemporal correlation analysis unit includes a partitioned time series data integration module, a driving factor knowledge graph construction module, and a spatiotemporal evolution and driving analysis module;

[0088] The partitioned time-series data integration module receives the spatial distribution results of turbidity in the entire lake area, the results of turbidity inversion and adaptation to sub-partitions, and the meteorological and hydrological auxiliary datasets. Based on the partition boundaries of the turbidity inversion and adaptation to sub-partitions, it statistically analyzes the time-series turbidity data and corresponding meteorological and hydrological data of each sub-partition to generate a partitioned time-series dataset.

[0089] The knowledge graph construction module for driving factors receives partitioned time-series datasets, extracts entities and relationships of turbidity and meteorological and hydrological elements, and constructs a knowledge graph of lake turbidity driving factors.

[0090] The spatiotemporal evolution and driving analysis module receives a knowledge graph of lake turbidity driving factors, mines driving factors that meet preset contribution thresholds in the turbidity changes of each sub-region, analyzes the spatiotemporal evolution law of turbidity, and obtains the turbidity change trend and driving factor analysis results.

[0091] like Figure 2 As shown, a deep learning-driven remote sensing identification and analysis method for lake water turbidity includes the following steps:

[0092] Step 1: The multi-source raw data acquisition module acquires multi-source satellite remote sensing images of the lake, measured water turbidity data, and meteorological and hydrological auxiliary data. The acquired raw data is preprocessed to obtain water spectral dataset, measured turbidity sample dataset, and meteorological and hydrological auxiliary dataset.

[0093] Step 11: The multi-source raw data acquisition module acquires three types of basic data;

[0094] The first type of data consists of multi-source satellite remote sensing images, specifically Landsat-8 / 9 OLI, Sentinel-2MSI, Gaofen-1, or Gaofen-2 multispectral satellite images. The image data is stored in the standard GeoTIFF format, and the corresponding metadata file records observation geometric parameters such as imaging time, solar elevation angle, and satellite zenith angle.

[0095] The second type of data consists of measured turbidity data of lake water, obtained hourly by automatic lake monitoring stations. Each record includes at least the station's latitude and longitude coordinates, the observation date and time, and the measured turbidity value, with the unit of turbidity being NTU.

[0096] The third type of data is meteorological and hydrological auxiliary data, which is obtained hourly by automatic lake observation stations. It includes observation records of elements such as wind speed, wind direction, precipitation, temperature, and water level, and is stored in time series. Each record is accompanied by station information including latitude and longitude coordinates and a timestamp.

[0097] Step 12: The remote sensing image preprocessing module performs radiometric correction, atmospheric correction, water body region extraction, and water body spectral data extraction sequentially on the received multi-source satellite remote sensing images. Since the data distribution agency has already performed geometric correction on the satellite images, no further geometric correction operation is performed.

[0098] Step 121, Radiometric correction of satellite imagery: Using the radiometric calibration coefficients inherent in each satellite imagery, the digital quantization (DN) values ​​of the imagery are converted into physically meaningful apparent radiance or top atmospheric reflectance.

[0099] Step 122, Atmospheric correction of satellite imagery: Using an atmospheric radiative transfer model and combining meteorological data, the apparent radiance of the satellite imagery is further converted into surface reflectance, thereby eliminating the interference of atmospheric molecular scattering, aerosol scattering, and absorption effects on the water body reflection signal, and completing the atmospheric correction of the satellite imagery.

[0100] Step 123, Lake Water Area Extraction from Satellite Imagery: To remove non-lake features such as land from the satellite imagery and retain only lake water pixels, water area extraction processing is performed on the corrected satellite imagery. Water area extraction can employ the normalized water index method, which calculates the normalized ratio of each pixel in the green band to the near-infrared band, applies a threshold to the calculation results, and identifies pixels exceeding a preset threshold as water pixels.

[0101] Step 13: The multi-source data spatiotemporal registration module sets the geographic coordinates of satellite imagery and lake automatic observation stations as a unified spatial coordinate system and matches the time window. It then completes the spatiotemporal matching and standardization of the initial water spectral data of satellite imagery, the measured turbidity ground data of lakes, and the meteorological and hydrological auxiliary data, resulting in a water spectral dataset, a measured turbidity dataset, and a meteorological and hydrological auxiliary dataset that are consistent with the observation time and geographical location.

[0102] Specifically, during spatial matching, the multi-source data spatiotemporal registration module unifies the geographic coordinates of satellite imagery and automatic lake observation stations to the same spatial coordinate system, such as using the WGS-84 coordinate system and UTM projection. Then, based on the latitude and longitude coordinates of the automatic lake observation stations, it extracts the reflectance values ​​of the corresponding locations of all bands of lake observation stations from the satellite imagery to obtain initial water spectral sample data, which is then matched with measured turbidity and meteorological / hydrological data. During temporal matching, the multi-source data spatiotemporal registration module uses the satellite's transit time for lake imaging as a benchmark, setting a time matching window of ±0.5 hours with the automatic lake observation station's observations. It matches the turbidity data and meteorological / hydrological data from the automatic lake observation stations within half an hour of the satellite transit time with the water spectral data from the satellite imagery.

[0103] After completing the spatiotemporal matching of the data, the multi-source data spatiotemporal registration module standardizes the three types of data, unifies the data format and missing value identification method, fills missing values ​​in the samples with the neighborhood mean, and adds necessary data quality labels. Finally, it forms a complete water body spectral dataset, measured turbidity dataset, and meteorological and hydrological auxiliary dataset, which are respectively input into the water body optical type partitioning unit, the partition adaptation deep learning turbidity inversion unit, and the turbidity spatiotemporal correlation analysis unit.

[0104] Step 2: Input the water body spectral dataset into the water body optical type partitioning unit, and use spectral feature extraction and cluster analysis methods to classify the optical type of the lake satellite image to obtain the initial water body optical parent partitioning result;

[0105] Step 21: The water body spectral feature extraction module processes the water body spectral dataset, extracting reflectance of characteristic bands, constructing ratios of each band, and extracting peak and valley features of the water body spectrum as spectral feature quantities reflecting the optical properties of the water body, generating a water body spectral feature set, which provides a feature space for subsequent clustering analysis of lake satellite imagery. Specifically:

[0106] Extracting reflectance features: Selecting reflectance values ​​of blue, green, red, and near-infrared bands that are sensitive to changes in water composition directly from the water spectral dataset. The reflectance of these bands directly reflects the combined effect of optically active substances such as phytoplankton chlorophyll, suspended matter, and colored soluble organic matter in the water, and is a basic feature for distinguishing different optical types of water.

[0107] Extracting band ratio features: By performing ratio calculations on the reflectance of different bands, multiple band ratios are obtained as feature quantities, such as the ratio of blue band reflectance to green band reflectance, the ratio of green band reflectance to red band reflectance, and the ratio of near-infrared band reflectance to red band reflectance. Band ratios can, to a certain extent, eliminate the influence of atmospheric correction residual errors and observation geometric changes on the absolute value of reflectance, highlighting the relative spectral differences between water bodies of different optical types.

[0108] Extracting spectral peak and valley features: By analyzing the morphology of water spectra, characteristic quantities reflecting spectral morphology are extracted. These include the presence of relative reflection peaks or valleys within a specified band combination range, the band positions corresponding to the reflection peaks or valleys, the relative peak height of the reflection peaks, and the relative valley depth of the reflection valleys. These spectral peak and valley features are closely related to the absorption and scattering characteristics of phytoplankton pigments and suspended particulate matter in the water, as well as the absorption characteristics of soluble organic matter, and can provide richer information on the optical properties of water bodies than single-band reflectance.

[0109] Finally, the water body spectral feature extraction module extracts and combines the above three types of features to form a water body spectral feature set, which constitutes the feature vector for satellite image clustering.

[0110] Step 22: The water body optical clustering module receives the water body spectral feature set as a training sample for lake satellite image clustering. It then uses clustering analysis to perform clustering processing on the lake satellite image, dividing the entire lake area into several types of regions with different optical properties.

[0111] Specifically, clustering analysis methods can include commonly used classification methods in this field, such as K-means clustering, ISODATA self-organizing clustering, or Gaussian mixture model clustering. During the clustering process, feature vectors from the water body spectral feature set are used as input samples. By iteratively optimizing the objective function, samples that are close in distance in the feature space are grouped into one class. This ensures that water body pixels belonging to the same optical type have high similarity in spectral features, while water body pixels of different optical types exhibit significant spectral differences.

[0112] Furthermore, the number of clusters can be preset by the user based on the actual complexity of the lake water, or it can be adaptively determined using clustering effectiveness evaluation metrics. Commonly used clustering effectiveness evaluation metrics include the silhouette coefficient and the Davies-Bouldin index. By comparing the evaluation metric values ​​under different numbers of clusters, the number of clusters that achieves a better evaluation metric level is selected as the final number of clusters.

[0113] After clustering, each water cell in the lake satellite image is assigned a category label, indicating the optical type to which the cell belongs. A set of spatially adjacent cells with the same category label constitutes a preliminary water body optical partition. All water body optical partitions together form the preliminary water body optical partitioning result, which is then input into the parent partition post-processing module for further processing.

[0114] Step 23: The parent partitioning post-processing module performs spatial connectivity regularization on the partition boundaries and merges patches with areas smaller than a preset threshold to obtain the initial water body optical parent partitioning results. To address potential issues such as irregular partition boundaries and scattered small patches that may arise after clustering, the parent partitioning post-processing module performs the following two steps:

[0115] Step 231: Perform spatial connectivity regularization processing on the parent partition. In the partitioning results generated by the clustering algorithm, pixel sets belonging to the same optical type but spatially separated by other types often appear as multiple disconnected independent patches. The parent partitioning post-processing module performs connectivity analysis on the pixel sets of each optical type using the eight-neighborhood connectivity criterion to identify each independent connected component. For connected components with larger areas and more regular boundaries, they are directly retained as valid partitions; for jagged or irregular protrusions in the boundaries caused by clustering boundary jitter, morphological smoothing or boundary refinement operations based on spatial neighborhood relationships can be used to make the partition boundaries more regular and smooth.

[0116] Step 232 involves merging patches with areas smaller than a preset threshold. During clustering, some small, scattered patches inevitably occur. These patches typically contain only a few pixels and may be due to noise interference or transitional pixels at the boundaries. They lack independent optical type representativeness and are detrimental to the stable construction of the subsequent turbidity inversion model. The parent partition post-processing module calculates the area of ​​each connected component and re-merges patches with areas smaller than the preset area threshold into the spatially adjacent optical type partition with the largest area. The preset area threshold can be determined based on the spatial resolution of the remote sensing image and the size of the lake area; for example, it can be set to several hundred to several thousand square meters.

[0117] After processing in steps 231 and 232, the resulting partitioning is the initial optical parent partitioning result for the water body. In this initial optical parent partitioning result, each partition has good spatial continuity and regular boundary morphology, providing a reasonable initial optical partitioning framework for subsequent partitioning adaptation to deep learning turbidity inversion units.

[0118] Based on the initial water body optical parent partitioning results, the water body spectral dataset and measured turbidity sample dataset within each partition are input into the partition-adaptive deep learning turbidity inversion unit to perform iterative optimization of the partition boundary and the turbidity inversion model.

[0119] Step 3: Using a partition-adapted deep learning turbidity inversion unit, a local iterative clustering collaborative inversion method driven by cross-validation sample residuals is employed to perform linked iterative optimization of the partition boundaries and turbidity inversion models within each water body's optical parent partition. This yields the turbidity inversion adaptation sub-partition results and the spatial distribution of turbidity across the entire lake area. The specific implementation process is as follows: Figure 3 As shown;

[0120] Step 31: The cross-validation sample residual generation module first groups the spatiotemporally matched measured turbidity sample dataset and the water body spectral dataset according to the partition boundaries given in the initial water body optical parent partition results, forming sample subsets corresponding to each water body optical parent partition; then, turbidity remote sensing modeling and residual calculation are performed on the sample subsets of each partition, the specific process of which is as follows:

[0121] Step 311: The measured turbidity sample dataset and water spectral data are grouped according to the initial water body optical parent partition, forming sample subsets corresponding to each water body optical parent partition. Specifically, for the measured turbidity sample points of the lake automatic observation station, the cross-validation sample residual generation module determines which initial water body optical parent partition the sample point falls into based on its latitude and longitude coordinates. If the sample point happens to fall on the partition boundary line, it is assigned to the larger area of ​​the adjacent partition according to the area dominance principle, and its spectral vector and measured turbidity value are assigned to the sample subset of that partition. Thus, each water body optical parent partition in the initial water body optical parent partition result obtains a set of exclusive sample subsets, which contain the spectral data and corresponding measured turbidity values ​​of all measured sample points within the partition.

[0122] Step 312: For each partition's sample subset, a 5-fold cross-validation method is used to divide it into training and validation groups. Specifically, the sample subset corresponding to each optical parent partition is randomly divided into 5 non-overlapping subsets. Each time, 4 subsets are selected as the training group, and the remaining subset is selected as the validation group, for a total of 5 rounds of training and validation. In each round, the spectral vector of the training group samples is used as input, and the measured turbidity value is used as the target output to train a turbidity inversion model specific to that partition.

[0123] In this embodiment, the turbidity inversion model adopts a multilayer perceptron structure. During pre-modeling experiments, it was found that spectral peak and valley features had very limited impact on improving the accuracy of turbidity inversion. To control model complexity and training stability, this embodiment only uses single-band reflectance and band ratio features as model inputs. Specifically, the turbidity inversion model input layer has 6 nodes, corresponding to 6 spectral feature values: blue band reflectance, green band reflectance, red band reflectance, near-infrared band reflectance, the ratio of green to blue band reflectance, and the ratio of near-infrared to red band reflectance. The turbidity inversion model has two hidden layers: the first hidden layer has 64 nodes, and the second hidden layer has 32 nodes, both using the ReLU activation function. The output layer has 1 node, directly outputting the predicted turbidity value without an activation function. During training, mean squared error was used as the loss function, the Adam optimizer was selected, the learning rate was set to 0.001, the batch size was set to 16, the maximum number of training rounds was set to 200, and an early stopping mechanism was set to terminate training early when the loss did not decrease for 10 consecutive rounds.

[0124] Step 313: Calculate the predicted turbidity value of the corresponding validation group sample through the turbidity inversion model of each partition, and obtain the inversion residual of each sample by combining the corresponding measured turbidity value, forming a cross-validation inversion residual point set with spatial location and spectral attributes.

[0125] In each round of cross-validation, the cross-validation residual generation module uses the trained partitioned turbidity inversion model to predict the turbidity of each sample in the validation group, obtaining the predicted turbidity value. For the first sample in the validation group... For each sample, the inversion residual is calculated using the following formula:

[0126] ;

[0127] In the formula, Indicates the first The inversion residuals of each sample This represents the measured turbidity value of the sample. This represents the predicted turbidity value for the sample. A positive residual indicates that the model underestimates the actual turbidity; a negative residual indicates that the model overestimates the actual turbidity. After five rounds of cross-validation, each sample is used as a validation sample exactly once, yielding a corresponding inversion residual value. The residuals of all samples within all optical parent partitions are summed to obtain the cross-validation inversion residual sample point set. In this sample set, each residual sample corresponds to a measured turbidity value from an automatic lake station, along with the residual value, the latitude and longitude coordinates of the sample point, the initial optical parent partition number, and the corresponding spectral vector.

[0128] Finally, the cross-validation residual generation module inputs the generated cross-validation inversion residual sample point set into the partition attribution sample residual screening module.

[0129] Step 32: The partition attribution sample residual screening module screens the cross-validation inversion residual sample point set to obtain candidate partition attribution residual samples.

[0130] The partition attribution sample residual screening module receives the cross-validation inversion residual sample point set and removes error components not caused by partitioning to obtain candidate partition attribution residual samples. Since the sources of residuals in the cross-validation inversion residual point set are not singular, some residuals are caused by unreasonable water body optical partition boundaries or misclassification of samples, while others are caused by factors unrelated to partitioning, such as instrument measurement noise, atmospheric correction residual errors, or model training randomness. The purpose of the partition attribution sample residual screening module is to eliminate the latter type of residuals as much as possible and retain the former type of residual signals that truly point to partition boundary problems. The specific screening process is as follows:

[0131] Step 321: Perform error standardization on the cross-validation inversion residual sample point set within each water body optical parent partition to unify the dimensional scale of the residuals.

[0132] The optical properties and turbidity variation ranges of different water body optical parent zones often differ, and the original residuals generated by each zone may also vary significantly in amplitude. Direct cross-zone comparisons or mixed processing will introduce bias. Therefore, the zone attribution sample residual screening module standardizes the residuals within each water body optical parent zone separately. For the first... For each water body optical parent partition, first calculate the mean of the residual values ​​of all residual points within that partition. and standard deviation Then, the original residual value of each residual point within that partition. Perform Z-score standardization; if the standard deviation of the residuals within this partition... If the value is 0, then the standardized residual values ​​of all samples in that partition are uniformly set to 0; calculated using the following formula:

[0133] ;

[0134] In the formula, This represents the standardized residual value, which is dimensionless. This represents the original residual value, in NTU. Indicates the first The mean of the residual values ​​of all residual points within a single optical parent partition, in NTU; Indicates the first The standard deviation of the residual values ​​of all residual points within a water body optical parent partition is expressed in NTU.

[0135] After standardization, the residuals of each partition are converted into dimensionless quantities with a mean of 0 and a standard deviation of 1.

[0136] Step 322: Calculate the common offset residual components of the residuals within the optical parent partition of each water body, remove the corresponding common offset residual components from each sample residual, and obtain the residual point set after system bias correction.

[0137] The common offset residual component refers to the systematic offset that almost all residual points share within a parent optical partition of a water body. This offset typically originates from overall factors such as atmospheric correction residual bias and sensor calibration system errors, rather than problems inherent in the partition boundaries themselves. For the first... For each water body optical parent partition, the partition attribution sample residual screening module takes the median of the standardized residuals of all residual points within that partition. This median is used as an estimate of the common offset residual components. The median is then subtracted from the standardized residual values ​​at each residual point to obtain the residual values ​​after systematic bias correction. :

[0138] ;

[0139] In the formula, This represents the corrected residual value, which is dimensionless. Represents the standardized residual value; Indicates the first The median of the standardized residuals of all residual points within a parent optical partition of a water body. The corrected residual value reflects the individual bias caused by the inconsistency between the spectral characteristics of each residual point and the mainstream characteristics of its partition.

[0140] Step 323: Identify and remove isolated residual samples from the corrected residual point set based on spatial adjacency relationship to obtain candidate partition attribution residual samples;

[0141] After steps 321 and 322, the corrected residual point set may still contain some isolated residual points caused by measurement noise or local random factors. However, deviations truly caused by unreasonable partition boundaries usually do not appear at a single isolated point, but rather in clusters within a certain spatial range on both sides of the boundary, forming continuous clusters of residuals with the same sign. The partition attribution sample residual filtering module utilizes this characteristic to perform spatial filtering on the corrected residual point set. Specifically, for each residual point, with its spatial location as the center, within a preset spatial neighborhood radius... Within the range, excluding the residual point itself, count the number of neighboring residual points with the same sign as its residual. Spatial neighborhood radius The spatial resolution is determined based on the spatial resolution of the remote sensing image, typically 30 to 50 times the image's spatial resolution. In this example, the spatial resolution is 30m for a Landsat-8 image. Take 1200m. If the number of neighboring residual points... Below the preset neighborhood density threshold (In this embodiment) If a value of 3 is selected, the residual point is considered an isolated point and is discarded. The residual points retained after spatial filtering constitute the candidate partition attribution residual point set. Each residual point in this candidate partition attribution residual point set corresponds to a ground-measured turbidity sample point, carrying the corrected residual value, latitude and longitude coordinates, the optical parent partition number, and the spectral vector. These residual points are spatially clustered and have consistent residual signs, which can clearly indicate the locations where partition boundaries may need to be adjusted.

[0142] Finally, the partition attribution sample residual screening module inputs the candidate partition attribution residual point set into the local clustering boundary adjustment module.

[0143] Step 33: The local clustering boundary adjustment module processes the candidate partition attribution residual samples and the initial water body optical parent partition results, and generates local partition adjustment candidates within the corresponding optical parent partition.

[0144] The local clustering boundary adjustment module includes a residual connected component identification submodule and a partition candidate construction submodule, wherein:

[0145] The residual connected component identification submodule receives candidate partition attribution residual samples and initial water body optical parent partition results. Based on spatial adjacency, it identifies continuous clustered regions with consistent residual signs, obtains residual connected components, and inputs them into the partition candidate construction submodule. Specifically, the residual connected component identification submodule takes the spatial coordinates of each residual point in the candidate partition attribution residual point set as input. According to a preset spatial adjacency rule, it groups adjacent residual points with the same residual sign into the same connected component. The spatial adjacency rule adopts a density connectivity criterion based on a preset distance threshold; that is, when the spatial straight-line distance between two residual points does not exceed the preset adjacency distance threshold, they are determined to belong to the same connected component. This adjacency distance threshold is related to the spatial neighborhood radius in step 323. Take the same value. After the connected component analysis is completed, the residual connected component identification submodule outputs several residual connected components. Each residual connected component is a set of residual points with the same residual sign and continuous spatial distribution, indicating that the water pixels in the area covered by the residual point set may not be reasonably assigned in the current water optical parent partition. Therefore, these residual connected components are then sent to the partition candidate construction submodule.

[0146] The partition candidate construction submodule receives residual connected components and uses them as adjustment seeds to generate local partition adjustment candidates, including splitting and merging types, within the corresponding water body optical parent partition. The partition candidate construction submodule generates corresponding partition adjustment candidate schemes for each residual connected component and its corresponding water body optical parent partition, as follows:

[0147] Step 331: For the region corresponding to the residual connected component, extend the sub-region boundary within the optical parent region of the single water body to form a partition splitting candidate; and use the partition splitting candidate as a new sub-region.

[0148] For each residual connected component, the partition candidate construction submodule first determines which water body optical parent partition it falls into, and then marks the remote sensing image water body pixels corresponding to each residual point in that residual connected component as initial seed pixels, which together constitute the seed region. The growth range of the seed region is limited to the spatial range of the current water body optical parent partition and will not cross the boundary of the water body optical parent partition.

[0149] The seed region growth process is constrained by spectral feature similarity: starting from the boundary cell of the seed region, neighboring cells within the current water body optical parent partition are checked one by one using the eight-neighbor adjacency criterion. For the neighboring cells to be checked... Calculate its spectral eigenvector. With the mean vector of pixel spectral features within the seed region Euclidean distance between :

[0150] ;

[0151] In the formula, Indicates the pixel to be inspected The spectral feature vector is composed of reflectance and band ratio extracted by the water body spectral feature extraction module. This represents the mean vector of the spectral feature vectors of all pixels within the current seed region; This represents the L2 norm of a vector.

[0152] like Less than the preset spectral similarity threshold Then the adjacent pixel The newly included pixels are incorporated into the growth region, and the mean vector of spectral features of the seed region is immediately recalculated and updated using the newly included pixels before checking the next neighboring pixel. A spectral similarity threshold is used. Generally, the standard deviation of each spectral feature dimension of the seed region is used as a reference, usually 1.5 to 2.5 times the standard deviation, so as to ensure that the spectral features within the grown seed region remain highly consistent.

[0153] Repeat the seed region growth process described above until no more suitable neighboring pixels can be added within the current water body optical parent partition, or until the boundary of the water body optical parent partition is reached. The resulting region boundary is a partition splitting candidate, and it is recommended to separate this region from the original water body optical parent partition as a new sub-partition.

[0154] Step 332: For spatially adjacent sub-partitions within the same water body optical parent partition, generate partition merging candidates;

[0155] In the first iteration, since no sub-regions have yet formed within the parent optical region of each water body, no candidate regions for merging are generated. From the second iteration onwards, the candidate region construction submodule checks each spatially adjacent sub-region within the same parent optical region of the water body under the current partition state. If the spectral feature centers of two adjacent sub-regions are close in the current partition state, it indicates that their optical type differences may not be significant, and merging can be considered. For two adjacent sub-regions A and B, their spectral feature mean vectors are first calculated separately. and Then calculate the Euclidean distance between them. :

[0156] ;

[0157] In the formula, , These are the mean vectors of the spectral feature vectors of all pixels in sub-partitions A and B before the start of the current iteration round, respectively.

[0158] like Less than the preset merging distance threshold If the two adjacent sub-partitions are not found in the specified range, they will be merged into one partition, which will then be considered as a partition merge candidate. Merge distance threshold. The method can be determined by referring to the statistical distribution of the spectral center distances of all adjacent sub-regions in the entire lake. For example, take 0.5 times the median of the spectral center distances of all adjacent sub-regions. In this way, only those adjacent sub-regions with significantly smaller spectral differences will be included in the candidate for sub-region merging.

[0159] Step 333: Perform spectral feature separability verification on all partition splitting and partition merging candidates. The partition candidates that pass the verification are used as the final local partition adjustment candidates.

[0160] Partition adjustment is not only about reducing inversion errors, but also about ensuring that the adjusted partitions still have clear physical meaning in optics. Therefore, the partition candidate construction submodule needs to perform spectral feature separability verification on each partition splitting candidate and partition merging candidate. This implementation uses the JM distance as the separability metric. For the newly formed sub-partition and its original water body optical parent partition (the part remaining after deducting the newly formed sub-partition) in the partition splitting candidate, the mean vector and covariance matrix of their spectral feature vectors are calculated respectively. If the covariance matrix is ​​singular, a small regularization term is added to its diagonal to ensure matrix invertibility, and then the Bach distance is calculated according to the following formula. Distance from JM :

[0161] ;

[0162] ;

[0163] In the formula, , These represent the mean vectors of the spectral feature vectors of the two regions to be compared; , Let represent the covariance matrices of the spectral eigenvectors of the two regions, respectively. Represents the determinant of a matrix; Represents the transpose of a vector or matrix. This represents the JM distance, with a value between 0 and 2. The closer it is to 2, the easier it is to distinguish the two regions on the spectrum.

[0164] The partition candidate construction submodule sets two verification thresholds: a splitting separability threshold. and merge separability threshold Both satisfy For partition split candidates, when the calculated... If the newly formed sub-partition has sufficient spectral differences from the remaining part, and the split is reasonable, then the partition splitting candidate passes the verification. For partition merging candidates, when... If the spectral differences between the two sub-partitions are not significant to begin with, and the internal consistency can still be guaranteed after merging, then the candidate for merging the sub-partitions passes the verification. and The specific value can be selected through preliminary experiments based on the spectral characteristics of the actual lake water. In this embodiment, Take 1.0, Taking 0.5, both satisfy the condition. The relationship is logically consistent. Candidates that pass the spectral feature separability verification are used as final local partition adjustment candidates and sent to the partition inversion linkage iteration module.

[0165] Step 34: The partition inversion linkage iteration module receives local partition adjustment candidates, determines the validity of the partition adjustment, and controls the iteration process. After the iteration terminates, it outputs the turbidity inversion adaptation sub-partition results and the spatial distribution results of turbidity in the entire lake area. Before entering the iteration cycle, the cross-validation residual generation module needs to generate a cross-validation inversion residual point set based on the initial water body optical parent partition results. Then, the partition attribution sample residual screening module processes the residual points to obtain the candidate partition attribution residual point set. Finally, the local clustering boundary adjustment module generates the first batch of local partition adjustment candidates based on this residual point set and sends them to the partition inversion linkage iteration module for evaluation, thus starting the first round of iteration. The partition inversion linkage iteration module includes an adjustment evaluation sub-module and an iteration control sub-module, wherein:

[0166] The adjustment evaluation submodule receives local partition adjustment candidates, calculates the cross-validation inversion error and partition complexity before and after adjustment for the corresponding regions of the local partition adjustment candidates, generates adjustment evaluation index by combining accuracy gain and complexity cost, and inputs the adjustment evaluation results into the iterative control submodule.

[0167] Specifically, the adjustment evaluation submodule receives local partition adjustment candidates from the local clustering boundary adjustment module. For each candidate, it calculates the cross-validation inversion error and partition complexity before and after adjustment within the affected spatial region, and then generates an adjustment evaluation index. The affected region is defined as follows: if it is a partition splitting candidate, the affected region includes the newly formed sub-partition and the remaining part of the original water body optical parent partition after deducting the newly formed sub-partition; if it is a partition merging candidate, the affected region is the area covered by the merged new partition.

[0168] The calculation method for cross-validation inversion error is consistent with that of the cross-validation residual generation module. Cross-validation modeling and prediction are performed on measured turbidity sample points within the affected area under both the pre-adjustment and post-adjustment zoning schemes to calculate the root mean square error before and after adjustment. :

[0169] ;

[0170] In the formula, This indicates the number of measured turbidity sample points within the affected area; , The first The measured and predicted turbidity values ​​for each sample are given, both in NTU. Let the root mean square error before adjustment be... The adjusted version is The reduction in cross-validation inversion error for:

[0171] ;

[0172] A positive value indicates that the turbidity inversion error has decreased and the accuracy has improved after adjustment; the unit is NTU. When ΔE is negative, it indicates that the modeling parameters need to be optimized and adjusted.

[0173] Furthermore, regarding partitioning complexity, the adjustment evaluation submodule is measured by the change in the number of partition boundary cells within the affected area; where partitioning complexity is measured by the total number C of partition boundary cells within the affected area; let the number of partition boundary cells before adjustment be... The adjusted version is The increment of partitioning complexity for:

[0174] ;

[0175] When it is a positive integer, it means that the partition boundary has become longer and the partitioning scheme has become more complex. When the value is a negative integer, it indicates that the partition boundary has become shorter and the partitioning scheme has been simplified.

[0176] Furthermore, the evaluation submodule is adjusted by calculating the evaluation indicators using the following formula. :

[0177] ;

[0178] In the formula, It is a preset complexity cost coefficient used to make a trade-off between improved accuracy and increased complexity, and its unit is NTU / pixel. The value needs to be determined through experiments based on the specific application scenario, and is generally selected between 0.01 and 0.1. If the partitioning complexity is increased, the constraints on the increase in partitioning complexity will be stronger, and the optimization direction will be more inclined to maintain a simpler partitioning structure. If the target value is smaller, the optimization direction will focus more on pursuing higher inversion accuracy.

[0179] The effectiveness of the adjustment is determined based on the adjusted evaluation indicators. The judgment rule is straightforward: when the calculated... When this indicates that the accuracy gains from the adjustment outweigh the complexity costs, the adjustment evaluation submodule determines that the local partition adjustment candidate is a valid adjustment and adopts it; when If the accuracy gain is insufficient to compensate for the complexity cost, the local partition adjustment candidate is rejected.

[0180] Finally, the evaluation submodule outputs the results of the effectiveness determination of each local partition adjustment candidate to the iterative control submodule.

[0181] The iterative control submodule receives the adjustment validity judgment result, controls the iteration process to only update the sub-partitions affected by the partition adjustment and the turbidity inversion model, and outputs the turbidity inversion adaptation sub-partition results and the spatial distribution results of turbidity in the entire lake area after the iteration terminates.

[0182] Specifically, the iterative control submodule controls the advancement and cessation of the entire iterative optimization process based on the adjustment effectiveness judgment results. In each iteration, the iterative control submodule summarizes all local partition adjustment candidates generated in this round and their adjustment partition effectiveness judgment results, and applies them according to the adjustment evaluation indicators. The values ​​are sorted from high to low, and the candidates for local partition adjustment that are deemed valid are executed one by one. The execution includes two parts: first, updating the water optical partition boundaries of the corresponding area; and second, retraining the turbidity inversion model of the area using the updated partition scheme.

[0183] To control computational load, the iterative control submodule updates only the affected regions in each iteration, without updating the lake-wide partitions and turbidity inversion models. The affected regions are defined as follows: partition splitting candidates correspond to the newly formed sub-partitions and the remaining parts of the original water body optical parent partition; partition merging candidates correspond to the newly formed partitions after merging. Specifically, after executing a partition splitting candidate, the iterative control submodule only retrains the dedicated turbidity inversion models for the newly formed sub-partitions and the remaining parts of the original water body optical parent partitions, and recalculates the cross-validation residual point sets within these two partitions; the boundaries of other unaffected parent partitions and their turbidity inversion models are retained as is. After executing a partition merging candidate, only the dedicated turbidity inversion model is retrained for the merged new partition, and the residual point set is updated.

[0184] After all effective sub-partition adjustments have been completed and local partition adjustment candidate data has been updated in the current iteration round, the iteration control submodule uses the updated partitioning scheme and residual point set as input for the next iteration. Only in the local regions where partition adjustments have occurred does the cross-validation residual generation module, partition attribution sample residual screening module, and local clustering boundary adjustment module enter a new round of processing; local regions unaffected by adjustments retain their original calculation results. The iteration continues in this manner until a certain iteration generates no more effective partition adjustment candidates (i.e., all candidates generated in this round are no longer valid). All values ​​are satisfied If the number of iterations reaches the preset maximum number of iterations (usually set to 10-20 rounds), the iteration control submodule determines that the optimization process has converged or terminated, and the iteration terminates.

[0185] After the iteration terminates, the iteration control submodule outputs the final zoning scheme as the turbidity inversion adaptation sub-zoning result. In the turbidity inversion adaptation sub-zoning result, the boundaries of each sub-zoning achieve optimal fit with their respective dedicated turbidity inversion models. Using the turbidity inversion models of each adapted sub-zoning, water turbidity is inverted from the remote sensing images within the corresponding sub-zoning. Through image mosaicking, a spatial distribution map of turbidity covering the entire lake area is obtained.

[0186] The turbidity results from the adaptive sub-region inversion and the spatial distribution results from the lake-wide turbidity inversion are input into the turbidity spatiotemporal correlation analysis unit for subsequent analysis of turbidity spatiotemporal evolution patterns and mining of driving factors.

[0187] Step 4: The turbidity spatiotemporal correlation analysis unit receives the spatial distribution results of turbidity inversion of the entire lake area, the turbidity inversion results of the adapted sub-regions, and the meteorological and hydrological auxiliary datasets. It uses knowledge graph construction and correlation factor mining methods to analyze and process the spatiotemporal evolution law and driving factors of lake turbidity, and obtain the turbidity change trend and driving factor analysis results.

[0188] Step 41: The partition time series data integration module generates time series datasets for each sub-partition based on the boundaries of the adapted sub-partitions, using the long-term turbidity data retrieved from each sub-partition and the corresponding meteorological and hydrological data. The partition time series datasets are then input into the driving factor knowledge graph construction module.

[0189] Specifically, the partitioned time-series data integration module performs partitioned statistical analysis on the spatial distribution data of turbidity for each time phase according to the boundaries of each sub-partition. It calculates the average turbidity of all water pixels within each sub-partition, representing the turbidity value of that sub-partition at that time phase. After processing all time phases sequentially, the turbidity time series for each sub-partition is obtained. Simultaneously, the meteorological and hydrological data obtained from automatic lake observation stations within each sub-partition are processed: the meteorological and hydrological data includes multiple elements such as wind speed, wind direction, precipitation, temperature, and water level. Using the satellite's overpass imaging time of the lake as a benchmark, the average of meteorological and hydrological element data observed by multiple stations within each sub-partition with matching times is taken, completing the matching of turbidity for each sub-partition with each meteorological and hydrological element data. Each sub-partition forms a partitioned time-series data record, which includes sub-partition number, timestamp, turbidity, wind speed, wind direction, precipitation, temperature, and water level. The time-series data records of all sub-partitions are summarized into a partitioned time-series dataset and input into the driving factor knowledge graph construction module.

[0190] Step 42: The knowledge graph construction module for driving factors receives the partitioned time series dataset, constructs a knowledge graph of lake turbidity driving factors based on the correlation between turbidity and meteorological and hydrological element data, and inputs the knowledge graph of lake turbidity driving factors into the spatiotemporal evolution and driving analysis module.

[0191] First, entity information extraction. The driving factor knowledge graph construction module identifies three types of entities from the partitioned time-series dataset: 1) Sub-partition entities: Each turbidity inversion adaptation sub-partition is treated as an independent entity, with attributes including sub-partition number, spatial range, area, and optical type. 2) Turbidity state entities: The turbidity values ​​of each sub-partition at each time phase are discretized into different turbidity state levels according to a preset grading standard, such as low turbidity, lower turbidity, medium turbidity, higher turbidity, and high turbidity. Each turbidity state entity includes a level label, a turbidity value range, and a corresponding timestamp. 3) Meteorological and hydrological event entities: Specific event types are identified from the time-series of meteorological and hydrological elements, such as heavy precipitation events, strong wind events, and sustained high temperature events. The identification rules can be set according to the thresholds and durations of meteorological and hydrological elements. For example, a 24-hour cumulative precipitation exceeding a preset threshold is identified as a heavy precipitation event, and a wind speed exceeding a preset threshold and lasting for a certain duration is identified as a strong wind event.

[0192] Secondly, the construction of associations. The knowledge graph construction module for driving factors establishes two types of associations: the first is spatial attribution, connecting sub-region entities with turbidity state entities, indicating that a certain turbidity state occurs within a certain sub-region. The second is temporal driving relationships, connecting meteorological and hydrological event entities with turbidity state entities, indicating that a certain meteorological and hydrological event occurs before a certain turbidity state. The construction of temporal driving relationships requires setting a reasonable time lag window, for example, examining the changes in turbidity state of the corresponding sub-region within 1 to 7 days after a meteorological and hydrological event, to capture the lag effect of meteorological and hydrological factors on lake turbidity. The strength of the association can be determined by statistically analyzing the correlation between meteorological and hydrological events and changes in turbidity state under different time lag windows.

[0193] Finally, a knowledge graph of lake turbidity driving factors is generated. The knowledge graph construction module organizes and stores the extracted entities and relationships in a graph structure, with nodes representing entities and edges representing relationships. Both nodes and edges are equipped with attribute information to form a knowledge graph of lake turbidity driving factors, which is then input into the spatiotemporal evolution and driving analysis module.

[0194] Step 43: The spatiotemporal evolution and driving analysis module receives the knowledge graph of lake turbidity driving factors, mines the driving factors in each sub-region whose turbidity changes meet the preset contribution threshold (i.e., the confidence threshold of the association rule), analyzes the spatiotemporal evolution law of turbidity, and obtains the turbidity change trend and driving factor analysis results.

[0195] Specifically, the spatiotemporal evolution and driving analysis module performs the following analysis based on the knowledge graph of lake turbidity driving factors:

[0196] First, the spatiotemporal evolution pattern of turbidity is analyzed. The spatiotemporal evolution and driving analysis module extracts turbidity state information of each sub-region at each time phase from the knowledge graph of lake turbidity driving factors, and analyzes the turbidity change trend of each sub-region along the time axis. The trend analysis can use methods such as linear trend fitting and Mann-Kendall trend test to determine whether the turbidity of each sub-region shows an increasing, decreasing, or no obvious trend within the statistical period. At the same time, by comparing the turbidity change trends of different sub-regions in the same period, the spatial differentiation characteristics of turbidity changes within the lake can be identified. For example, it can be identified which areas of the lake are concentrated in sub-regions with rapid turbidity increases, which areas are distributed in sub-regions with decreasing turbidity, and which areas have relatively stable turbidity, thereby obtaining the spatiotemporal evolution pattern of turbidity across the entire lake.

[0197] Second, the module mines driving factors. The spatiotemporal evolution and driving analysis module utilizes the temporal driving relationships between meteorological and hydrological events and turbidity states stored in the lake turbidity driving factor knowledge graph to quantitatively assess the contribution of each meteorological and hydrological element to the turbidity changes in each sub-region. The contribution level can be calculated based on association rule mining or causal inference methods. For example, taking association rule mining as an example, the spatiotemporal evolution and driving analysis module extracts all association rules between meteorological and hydrological events and turbidity state changes from the lake turbidity driving factor knowledge graph, calculates the confidence level of each association rule, and the confidence level represents the probability that the turbidity state of the corresponding sub-region will undergo a specific change under the condition of a certain meteorological and hydrological event. Meteorological and hydrological elements involved in association rules with confidence levels exceeding a preset contribution threshold are identified as the main driving factors of turbidity changes in that sub-region. The preset contribution threshold can be set according to actual analysis needs, for example, it can be set to 0.6.

[0198] For each sub-region, the spatiotemporal evolution and driving factor analysis module outputs a list of its turbidity change trends and main driving factors, which are then summarized to form the turbidity change trend and driving factor analysis results. These results can be presented in tabular or thematic map form. The table lists the sub-region's number, area, turbidity change trend direction and rate, main driving factors, and their contribution levels. The thematic map uses different colors or symbols on the lake zoning map to represent the turbidity change trend type and main driving factor category of each sub-region.

Claims

1. A deep learning-driven remote sensing identification and analysis system for lake water turbidity, characterized in that, include: The multi-source heterogeneous data preprocessing unit is used to acquire multi-source satellite remote sensing images, measured turbidity data of lake water bodies, and meteorological and hydrological auxiliary data. It performs radiometric correction, atmospheric correction, water body area extraction, and spatiotemporal registration of remote sensing images and measured data. It also performs standardized preprocessing on the acquired raw data to obtain water body spectral datasets, measured turbidity sample datasets, and meteorological and hydrological auxiliary datasets. The water body optical type partitioning unit receives water body spectral datasets and uses spectral feature extraction and cluster analysis methods to perform optical type partitioning on the entire lake area, obtaining initial water body optical parent partitioning results. It includes: a water body spectral feature extraction module, which receives the water body spectral dataset, extracts the ratio of single-band reflectance to combined bands, and generates a water body spectral feature set; a water body optical clustering module, which receives the water body spectral feature set, uses cluster analysis methods to perform optical type clustering on the entire lake area, and obtains preliminary water body optical partitioning results; and a parent partitioning result adjustment module, which receives the preliminary water body optical partitioning results, performs spatial connectivity regularization on the partition boundaries, and removes patches with areas smaller than a preset threshold, obtaining the initial water body optical parent partitioning results. The partition adaptation deep learning turbidity inversion unit is used to receive the initial water body optical parent partition results, water body spectral dataset, and measured turbidity sample dataset. It adopts a local iterative clustering collaborative inversion method driven by cross-validation sample residuals to perform linkage iterative optimization processing on the partition boundaries and turbidity inversion model within each optical parent partition, so as to obtain the turbidity inversion adaptation sub-partition results and the spatial distribution results of turbidity in the entire lake area. The turbidity spatiotemporal correlation analysis unit is used to receive the spatial distribution results of turbidity in the entire lake area, the results of turbidity inversion and adaptation sub-regions, and meteorological and hydrological auxiliary datasets. It uses knowledge graph and correlation factor mining methods to analyze and process the spatiotemporal evolution law and driving factors of lake turbidity, and obtain the turbidity change trend and driving factor analysis results.

2. The deep learning-driven remote sensing identification and analysis system for lake water turbidity according to claim 1, characterized in that, The partition-adaptive deep learning turbidity inversion unit includes a cross-validation sample residual generation module, a partition attribution sample residual screening module, a local clustering boundary adjustment module, and a partition inversion linkage iteration module. The cross-validation sample residual generation module receives the initial water body optical parent partition results, water body spectral dataset, and measured turbidity sample dataset. It organizes the measured turbidity sample dataset and the corresponding water body spectral data according to the initial water body optical parent partition and generates a cross-validation inversion residual sample point set. The partition attribution sample residual screening module receives the cross-validation inversion residual sample point set, removes error components that are not caused by partitioning, and obtains candidate partition attribution residual samples. The local clustering boundary adjustment module receives candidate partition attribution residual samples and initial water body optical parent partition results, and generates local partition adjustment candidates within the corresponding water body optical parent partition; The partition inversion linkage iteration module receives local partition adjustment candidates, determines the validity of partition adjustment and controls the iteration process. After the iteration terminates, it outputs the turbidity inversion results of the adapted sub-partitions and the spatial distribution results of turbidity inversion of the entire lake area.

3. The deep learning-driven remote sensing identification and analysis system for lake water turbidity according to claim 1, characterized in that, The turbidity spatiotemporal correlation analysis unit includes a partitioned time series data integration module, a driving factor knowledge graph construction module, and a spatiotemporal evolution and driving analysis module; The partitioned time-series data integration module receives the spatial distribution results of turbidity inversion of the entire lake area, the turbidity inversion results of the adapted sub-partitions, and the meteorological and hydrological auxiliary datasets. According to the boundaries of the adapted sub-partitions, it integrates the time-series turbidity data of each sub-partition with the corresponding meteorological and hydrological data to generate the partitioned time-series dataset. The knowledge graph construction module for driving factors receives partitioned time-series datasets, establishes the association between turbidity and meteorological and hydrological element entities, and constructs a knowledge graph of lake turbidity driving factors. The spatiotemporal evolution and driving analysis module receives a knowledge graph of lake turbidity driving factors, mines driving factors that meet preset contribution thresholds in the turbidity changes of each sub-region, analyzes the spatiotemporal evolution law of turbidity, and obtains the turbidity change trend and driving factor analysis results.

4. A deep learning-driven remote sensing identification and analysis method for lake water turbidity, based on the analysis system described in any one of claims 1-3, to analyze the trend of turbidity changes in lake water, characterized in that... The steps include the following: S1: The multi-source raw data acquisition module acquires multi-source satellite remote sensing images, measured turbidity data of lake water, and meteorological and hydrological auxiliary data, and preprocesses the acquired raw data to obtain water spectral dataset, measured turbidity sample dataset, and meteorological and hydrological auxiliary dataset. S2, the reflectance, band ratio, and spectral peak-valley characteristics of turbidity-sensitive characteristic bands in the water body optical type partitioning unit are extracted from the water body spectral data, and the reflectance, band ratio, and spectral peak-valley characteristics are combined to form a water body spectral feature set; then, cluster analysis is performed on the water body spectral feature set, and each water body pixel in the entire lake area is assigned a category label to represent the optical type to which the pixel belongs; A set of pixels with the same category label and spatially adjacent constitutes a preliminary water body optical partition, and all water body optical partitions together form the preliminary water body optical partition result. Next, spatial connectivity regularization is performed on the partition boundaries, and patches with an area smaller than a preset threshold are merged to obtain the initial water body optical parent partition results. S3, using a cross-validation residual-driven local iterative clustering collaborative inversion method, the partition boundary of the water body optical parent partition and the turbidity inversion model are linked and iteratively optimized to obtain the turbidity inversion results of the adapted sub-partitions and the spatial distribution results of turbidity in the entire lake area. S4, the turbidity spatiotemporal correlation analysis unit, adopts knowledge graph construction and correlation factor mining methods. Based on the turbidity inversion results of the adaptive sub-region, the spatial distribution results of turbidity in the whole lake area, and the meteorological and hydrological auxiliary dataset, it analyzes and processes the spatiotemporal evolution law and driving factors of lake turbidity, and obtains the turbidity change trend and driving factor analysis results.

5. The deep learning-driven remote sensing identification and analysis method for lake water turbidity according to claim 4, characterized in that, In step S2, the steps for spatial connectivity regularization of the partition boundaries include: the parent partition result regularization module performs connectivity analysis on the pixel set of each optical type using the eight-neighbor connectivity criterion to identify each independent connected component; for connected components with larger areas and more regular boundaries, they are directly retained as valid partitions; for jagged or irregular protrusions on the boundaries caused by cluster boundary jitter, morphological smoothing or boundary refinement based on spatial neighborhood relationships is used to make the partition boundaries more regular and smooth. The steps for merging patches with an area smaller than a preset threshold include: the parent partition result normalization module calculates the area of ​​each connected component and re-merges patches with an area smaller than the preset area threshold into the optical type partition with the largest spatially adjacent area; wherein, the preset area threshold is determined based on the spatial resolution of the remote sensing image and the size of the lake area.

6. The deep learning-driven remote sensing identification and analysis method for lake water turbidity according to claim 4, characterized in that, Step S3, which involves obtaining the turbidity inversion results for the adaptor sub-region and the spatial distribution results of turbidity across the entire lake, includes the following steps: S31, the cross-validation sample residual generation module uses a cross-validation residual-driven local iterative clustering collaborative inversion method to generate a set of cross-validation inversion residual sample points, including: S311, group the measured turbidity sample dataset and the corresponding water body spectral data according to the initial water body optical parent partition, forming the sample subset corresponding to each water body optical parent partition; S312, For each sample subset, cross-validation is used to divide the training group and the validation group, and the turbidity inversion model of the corresponding partition is constructed based on the training group data; S313, calculate the predicted turbidity value of the corresponding validation group sample through the turbidity inversion model of each partition, and obtain the inversion residual of each sample by combining the corresponding measured turbidity value, forming a cross-validation inversion residual point set with spatial location and spectral attributes; S32, the partition attribution sample residual screening module filters the cross-validation inversion residual sample point set to obtain candidate partition attribution residual samples; including: S321, perform error standardization on the cross-validation inversion residual point set within each optical parent partition to unify the dimensional scale of the residuals; S322, count the common offset residual components of the residuals in each optical parent partition, remove the corresponding common offset residual components from each sample residual, and obtain the residual point set after system bias correction; S323, Spatial filtering is performed on the corrected residual point set to remove isolated residual samples from the corrected residual point set, and candidate partition attribution residual samples are obtained. S33, the local clustering boundary adjustment module processes the candidate partition attribution residual samples and the initial water body optical parent partition results, generating local partition adjustment candidates within the corresponding optical parent partition; including: S331, For the region corresponding to the residual connected component, extend the sub-partition boundary within the scope of the single optical parent partition to form a partition splitting candidate; and treat the partition splitting candidate as a new sub-partition. S332 generates partition merging candidates for spatially adjacent sub-partitions within the same water body optical parent partition; S333, perform spectral feature separability verification on all partition splitting and partition merging candidates, and use the candidates that pass the verification as the final local partition adjustment candidates; S34, the partition inversion linkage iteration module determines the effectiveness of local partition adjustment candidates, controls the iteration process based on the adjustment effectiveness determination result, and updates the sub-partitions affected by the partition adjustment and the turbidity inversion model; after the iteration terminates, it outputs the turbidity inversion results of the adapted sub-partitions and the spatial distribution results of turbidity in the whole lake area.

7. The deep learning-driven remote sensing identification and analysis method for lake water turbidity according to claim 6, characterized in that, The turbidity inversion model adopts a multilayer perceptron structure. The input layer has 6 nodes, corresponding to the reflectance of the blue light band, green light band, red light band, near-infrared light band, green light band to blue light band ratio, and near-infrared light band to red light band ratio, respectively. The hidden layer has 2 layers, with 64 nodes in the first hidden layer and 32 nodes in the second hidden layer. The ReLU function is used for both layers. The output layer has 1 node, which directly outputs the predicted turbidity value.

8. The deep learning-driven remote sensing identification and analysis method for lake water turbidity according to claim 6, characterized in that, In step S34, the cross-validation inversion error and partition complexity before and after adjustment are calculated for the corresponding regions of the local partition adjustment candidates. An adjustment evaluation index is then generated by combining the accuracy gain and complexity cost. The adjustment evaluation index satisfies the following relationship: ； In the formula, In order to adjust the evaluation indicators, The amount by which the cross-validation inversion error is reduced before and after partitioning adjustment. The increment in partition complexity before and after partition adjustment. This is a preset complexity cost coefficient; Next, based on the adjusted evaluation indicators, the results of the adjustment effectiveness determination are obtained: when When the local partition adjustment candidate is selected, it indicates that the adjustment is valid and should be adopted; when If this occurs, it indicates that the local partition adjustment candidate is an invalid adjustment and is rejected. Finally, based on the results of the adjustment effectiveness judgment, the iteration process is controlled and only the sub-partitions and turbidity inversion models affected by the partition adjustment are updated. After the iteration terminates, the final partition scheme is output as the turbidity inversion adaptation sub-partition result. Using the turbidity inversion model of each adaptation sub-partition, the spectral data of all water pixels in the corresponding sub-partition are inverted pixel by pixel to obtain the turbidity prediction value of each pixel, and a turbidity spatial distribution map covering the entire lake area is generated as the turbidity spatial distribution result of the entire lake area.

9. The deep learning-driven remote sensing identification and analysis method for lake water turbidity according to claim 4, characterized in that, Step S4, the specific steps for obtaining the turbidity change trend and driving factor analysis results, include: S41, the partitioned time-series data integration module integrates the time-series turbidity data and corresponding meteorological and hydrological data of each sub-partition according to the boundaries of the adapted sub-partitions, generating a partitioned time-series dataset; including: The final optimized spatial boundaries of each sub-partition are obtained from the turbidity inversion adaptation sub-partition results. Based on the spatial boundaries of each sub-region, the spatial distribution data of turbidity in each time phase is statistically analyzed by sub-region; the average value of the turbidity of all water pixels in each sub-region is taken as the turbidity value of that sub-region in that time phase; after processing all time phases in sequence, the time-series turbidity sequence of each sub-region is obtained. The meteorological and hydrological data from multiple automatic observation stations within each sub-region are averaged and correlated with the turbidity of the corresponding sub-region to obtain the meteorological and hydrological element value sequence of each sub-region. After completing the data association, the time-series turbidity sequence of each sub-partition is aligned and integrated with the meteorological and hydrological element sequence of the corresponding time period according to a unified time axis. After integration, each sub-partition forms a partition time-series data record, which includes the sub-partition number, timestamp sequence, turbidity time-series sequence, and time-series sequences of meteorological and hydrological elements; the partition time-series data records of all sub-partitions are summarized to form a partition time-series dataset. S42, Based on the partitioned time-series dataset, extract entities related to lake turbidity changes and the relationships between these entities, organize and store them according to a graph structure, and construct a knowledge graph of lake turbidity driving factors; details are as follows: Three types of entities are identified: sub-region entities, where each turbidity inversion adaptation sub-region is treated as an independent entity, with attributes including sub-region number, spatial extent, area, and optical type; turbidity state entities, where the turbidity values ​​of each sub-region at each time phase are discretized into different turbidity state levels according to a preset classification standard, and each turbidity state entity includes a level label, turbidity value range, and corresponding timestamp; and meteorological and hydrological event entities, which identify specific event types from the time series of meteorological and hydrological elements. Two types of relationships are constructed: spatial affiliation relationship, which connects sub-partition entities and turbidity state entities, indicating that a certain turbidity state occurs within a certain sub-partition; and temporal driving relationship, which connects meteorological and hydrological event entities and turbidity state entities, indicating that a certain meteorological and hydrological event occurs before a certain turbidity state in time. Generate a knowledge graph of lake turbidity driving factors: nodes represent entities, edges represent relationships, and both nodes and edges are attached with attribute information; S43, Based on the knowledge graph of lake turbidity driving factors, mine the driving factors that satisfy the preset contribution threshold for turbidity changes in each sub-region, analyze the spatiotemporal evolution law of turbidity, and obtain the turbidity change trend and driving factor analysis results; including: Analysis of the spatiotemporal evolution of turbidity: The spatiotemporal evolution and driving analysis module extracts the turbidity status information of each sub-region at each time phase from the knowledge graph of lake turbidity driving factors, and analyzes the turbidity change trend of each sub-region along the time axis; at the same time, by comparing the turbidity change trends of different sub-regions in the same period, the spatial differentiation characteristics of turbidity change within the lake are identified. Driving Factor Mining: The Spatiotemporal Evolution and Driving Analysis module utilizes the temporal driving relationship between meteorological and hydrological events and turbidity status stored in the knowledge graph of lake turbidity driving factors to quantitatively assess the contribution of each meteorological and hydrological element to the turbidity changes in each sub-region; and determines the main driving factors of turbidity changes in each sub-region based on the contribution level. For each sub-partition, a list of turbidity change trends and main driving factors is compiled to form the turbidity change trend and driving factor analysis results.

10. The deep learning-driven remote sensing identification and analysis method for lake water turbidity according to claim 9, characterized in that, The results of the analysis of turbidity change trends and driving factors are presented in the form of tables or thematic maps: the tables list the number, area, direction and rate of turbidity change trend, main driving factors and their contribution to each sub-region; the thematic maps use different colors or symbols on the lake zoning map to represent the type of turbidity change trend and the category of main driving factors for each sub-region.

Images