Water quality prediction system and method based on deep fusion and deep learning

Abstract

The application discloses a water quality prediction system and method based on deep fusion and deep learning, and particularly relates to the technical field of water quality prediction; the method is characterized by the following steps: acquiring multi-depth water temperature, dissolved oxygen and chlorophyll a data, as well as inflow and outflow of reservoir and satellite remote sensing image sequences, identifying the thermocline and oxygen consumption layer boundary based on Bayesian variable point detection, constructing a water column layered boundary table, and combining random replacement sequences with lag window transfer entropy to analyze the influence mechanism of multi-depth chlorophyll a on surface chlorophyll a; calculating the peak depth offset and water layer upturn reachable time to form an ectopic deep layer signal table, constructing positive and negative samples for deep learning network training, and finally outputting a high-precision surface chlorophyll a prediction map and a water quality early warning level. The application effectively fuses deep layer signals in water bodies and surface remote sensing data, and significantly improves the accuracy and real-time performance of water quality prediction.

Description

Technical Field

[0001] This invention relates to the field of water quality prediction technology, and more specifically, to a water quality prediction system and method based on deep fusion and deep learning. Background Technology

[0002] In the field of water quality early warning for water sources, especially in the short-term prediction of algal blooms, existing methods often employ a fusion strategy of in-situ sensor and satellite remote sensing data. Traditional data fusion models typically assume that the water quality of each layer has a consistent spatial mapping relationship. They extrapolate deep layer measurements to the water surface neighborhood through interpolation or statistical regression, and combine this with deep learning networks to learn time series patterns to generate early warning results. However, during actual reservoir thermostratification periods, the phenomenon of algae locally accumulating below the thermocline is very common, and there is no stable correlation between deep layer signals and surface conditions.

[0003] Existing fusion schemes fail to effectively address the physical discrepancies between profile information and surface data caused by vertical stratification of the water column when fusing monitoring data at different spatiotemporal scales. This makes it easy for the model to incorrectly map local deep high-value signals to the entire reservoir surface space under thermal stratification conditions, thereby inducing false early warnings that are unrelated to the actual development of algal blooms and reducing the accuracy and credibility of emergency dispatch decisions. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a water quality prediction system and method based on deep fusion and deep learning to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The water quality prediction method based on deep fusion and deep learning includes the following steps: S1: Acquire multi-depth water temperature data, multi-depth dissolved oxygen data, multi-depth chlorophyll a data, inflow and outflow data, and satellite surface chlorophyll a image sequence, and generate a multi-source spatiotemporal dataset according to the time grid; S2: Perform Bayesian variable point detection on multi-depth water temperature data and multi-depth dissolved oxygen data to identify thermocline boundaries and oxygen consumption boundaries, and generate a water column stratification boundary table; S3: Based on the water column stratification boundary table, the multi-depth chlorophyll a data is randomly rearranged within the time blocks corresponding to the water column stratification boundary of the same type to generate a stratified random substitution sequence. S4: Calculate the lag window transfer entropy from multi-depth chlorophyll a data and hierarchical random substitution sequence to satellite surface chlorophyll a image sequence, and compare and generate a depth surface information flow table; S5: When the information flow intensity of the surface information flow table is lower than the preset threshold, calculate the offset of the peak depth of chlorophyll a relative to the thermocline boundary, and estimate the upwelling time of the water layer based on the inflow and outflow data to form an ectopic deep signal table. S6: Construct positive and negative samples based on the deep surface information flow table and the ectopic deep signal table, train a deep learning network, and output a surface chlorophyll a prediction map and water quality warning level.

[0006] In a preferred embodiment, S1 specifically refers to: Collect multi-depth water temperature data, multi-depth dissolved oxygen data, and multi-depth chlorophyll a data at different monitoring depths and at each sampling time; Acquire inflow and outflow data and satellite surface chlorophyll a image sequences corresponding to each sampling time; Time registration was performed on each sampling and imaging time according to a unified time grid, and time markers, depth markers, and spatial location markers were established for multi-depth water temperature data, multi-depth dissolved oxygen data, multi-depth chlorophyll a data, inflow and outflow flow data, and satellite surface chlorophyll a image sequences to form a multi-source spatiotemporal dataset.

[0007] In a preferred embodiment, S2 specifically refers to: Based on the multi-depth water temperature data and multi-depth dissolved oxygen data corresponding to the same time marker, water temperature profiles and dissolved oxygen profiles are formed according to the depth markers; Bayesian variable point detection was performed on the water temperature profile along the depth direction to determine the locations of abrupt changes in the water temperature gradient as the boundaries of the thermocline. Bayesian variable point detection was performed on the dissolved oxygen profile along the depth direction to determine the locations of abrupt changes in dissolved oxygen concentration as oxygen consumption boundaries. A water column stratification boundary table is generated by associating the thermocline boundary and oxygen consumption boundary with the time marker.

[0008] In a preferred embodiment, S3 specifically refers to: Based on the water column stratification boundary table, the time blocks corresponding to the same type of water column stratification boundary are determined according to the correspondence between the thermocline boundary and the oxygen consumption boundary. Extract multi-depth chlorophyll a data corresponding to time markers within the time block, and form chlorophyll a sequences for each depth according to the depth markers; Within the time blocks corresponding to chlorophyll a sequences at each depth, the number of data, depth identifiers, and time block affiliations remain unchanged. The order of the time identifiers is randomly rearranged to generate a hierarchical random substitution sequence.

[0009] In a preferred embodiment, S4 specifically refers to: Based on time and spatial location identifiers, surface chlorophyll a sequences corresponding to chlorophyll a sequences at each depth are extracted from satellite surface chlorophyll a image sequences. Within the same time block, calculate the hysteresis window transfer entropy from each depth chlorophyll a sequence and the stratified random substitution sequence to the surface chlorophyll a sequence; By comparing the lag window transfer entropy with the depth identifier, the information flow intensity corresponding to the chlorophyll a sequence at each depth is determined, and a depth surface information flow table is generated.

[0010] In a preferred embodiment, S5 specifically refers to: Based on the deep surface information flow table, filter the time identifier and depth identifier corresponding to information flow intensity below a preset threshold; Multi-depth chlorophyll a data is extracted based on time markers to determine the peak depth of chlorophyll a. The thermocline boundary is extracted by combining the water column stratification boundary table, and the offset of the peak depth of chlorophyll a relative to the thermocline boundary is calculated. Based on the time marker associated with the inflow and outflow data, the time when the water layer can reach the top is estimated by combining the offset of the chlorophyll a peak depth relative to the thermocline boundary. An ex-situ deep-layer signal table is formed based on time markers, depth markers, offsets, and the time when the water layer can be upturned.

[0011] In a preferred embodiment, S6 specifically refers to: Based on the information flow table of depth and surface, filter the time identifier and depth identifier corresponding to the information flow intensity not lower than the preset threshold, and extract the corresponding multi-depth chlorophyll a data and surface chlorophyll a sequence to construct positive samples. Based on the ectopic deep signal table, the corresponding time marker, depth marker, offset, and water upturn reachable time are selected to extract the corresponding multi-depth chlorophyll a data, and negative samples are constructed by combining the offset and water upturn reachable time. The deep learning network is trained by using positive samples as surface prediction inputs and negative samples as mismatch constraint inputs. The multi-source spatiotemporal dataset is input into the trained deep learning network, which outputs a prediction map of surface chlorophyll a and a water quality warning level.

[0012] On the other hand, the present invention provides a water quality prediction system based on deep fusion and deep learning, comprising: Data acquisition module: acquires multi-depth water temperature data, multi-depth dissolved oxygen data, multi-depth chlorophyll a data, inflow and outflow flow data, and satellite surface chlorophyll a image sequence, and generates multi-source spatiotemporal datasets according to time grids; Layer identification module: Performs Bayesian variable point detection on multi-depth water temperature data and multi-depth dissolved oxygen data, identifies thermocline boundaries and oxygen consumption boundaries, and generates a water column layer boundary table; Sequence rearrangement module: Based on the water column stratification boundary table, it performs block random rearrangement on multi-depth chlorophyll a data within the time blocks corresponding to the water column stratification boundaries of the same type, generating a stratified random substitution sequence; Information flow table module: Calculates the lag window transfer entropy from multi-depth chlorophyll a data and hierarchical random substitution sequence to satellite surface chlorophyll a image sequence, and compares and generates a depth surface information flow table; Anomaly detection module: When the information flow intensity of the surface depth information flow table is lower than the preset threshold, calculate the offset of the chlorophyll a peak depth relative to the thermocline boundary, and estimate the upwelling time of the water layer based on the inflow and outflow flow data to form an anomaly deep layer signal table. Prediction and early warning module: Construct positive and negative samples based on the deep surface information flow table and the ectopic deep signal table, train a deep learning network, and output a surface chlorophyll a prediction map and water quality early warning level.

[0013] The technical effects and advantages of the water quality prediction system and method based on deep fusion and deep learning in this invention are as follows: By acquiring multi-depth water quality data, inflow and outflow data, and satellite surface chlorophyll a image sequences, and generating multi-source spatiotemporal datasets according to time grids, the temporal correspondence of data from different sources was improved. By identifying thermocline boundaries and oxygen consumption boundaries, a water column stratification boundary table was established, enabling the model to distinguish the physical states between different water layers. Stratified random substitution sequences were generated to provide a reference for judging the true correlation between deep chlorophyll a data and surface images. By calculating the hysteresis window transfer entropy and generating a depth-surface information flow table, it is possible to identify whether depth signals have information relationships that are transmitted to the surface. An ex-situ deep signal table was generated to avoid misjudging local deep high values ​​under thermal stratification conditions as surface algal bloom extension signals. By constructing positive and negative samples to train a deep learning network, the prediction results are made more consistent with the water column stratification state and the variation law of surface chlorophyll a, thereby improving the accuracy, stability, and reliability of surface chlorophyll a prediction maps and water quality warning levels. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the water quality prediction method based on deep fusion and deep learning of the present invention; Figure 2 This is a schematic diagram of the water quality prediction system based on deep fusion and deep learning according to the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. Example 1

[0016] Figure 1 The present invention provides a water quality prediction method based on deep fusion and deep learning, which includes the following steps: S1: Acquire multi-depth water temperature data, multi-depth dissolved oxygen data, multi-depth chlorophyll a data, inflow and outflow data, and satellite surface chlorophyll a image sequence, and generate a multi-source spatiotemporal dataset according to the time grid; S2: Perform Bayesian variable point detection on multi-depth water temperature data and multi-depth dissolved oxygen data to identify thermocline boundaries and oxygen consumption boundaries, and generate a water column stratification boundary table; S3: Based on the water column stratification boundary table, the multi-depth chlorophyll a data is randomly rearranged within the time blocks corresponding to the water column stratification boundary of the same type to generate a stratified random substitution sequence. S4: Calculate the lag window transfer entropy from multi-depth chlorophyll a data and hierarchical random substitution sequence to satellite surface chlorophyll a image sequence, and compare and generate a depth surface information flow table; S5: When the information flow intensity of the surface information flow table is lower than the preset threshold, calculate the offset of the peak depth of chlorophyll a relative to the thermocline boundary, and estimate the time when the water layer can be turned up based on the inflow and outflow data to form an ectopic deep signal table. S6: Construct positive and negative samples based on the deep surface information flow table and the ectopic deep signal table, train a deep learning network, and output a surface chlorophyll a prediction map and water quality warning level.

[0017] S1: Acquire multi-depth water temperature data, multi-depth dissolved oxygen data, multi-depth chlorophyll a data, inflow and outflow data, and satellite surface chlorophyll a image sequences, and generate a multi-source spatiotemporal dataset according to the time grid, including: A continuous, multi-scale monitoring scheme is established for the target reservoir to obtain comprehensive and accurate monitoring data. For example, multiple monitoring depths can be set, with data monitoring conducted at fixed locations 0.5 meters, 2 meters, 5 meters, 10 meters, 20 meters, and 30 meters below the water surface. Once the monitoring scheme is determined, based on the identified monitoring depths, in-situ multi-parameter water quality monitoring probes are used to collect multi-depth water temperature, dissolved oxygen, and chlorophyll a data at each sampling time. The in-situ multi-parameter water quality monitoring probes can simultaneously measure water temperature, dissolved oxygen, and chlorophyll a concentration, and through a built-in real-time synchronization function, the data from multiple monitoring depths can remain synchronized in time and space. For example, the probes can be set to automatically collect data every 30 or 60 minutes, recording the time information of each sampling moment, thereby obtaining a complete set of multi-depth water temperature, dissolved oxygen, and chlorophyll a data for each sampling time.

[0018] Simultaneously at each sampling moment, hydrological flow sensors installed at the reservoir's inlet and outlet spillway locations collect in-process flow data in real time. The flow data is acquired using flow meters, such as ultrasonic open channel flow meters or Doppler current meters, and the real-time flow data monitored at each sampling moment is recorded. The monitored flow data is expressed in cubic meters per second (m³ / s), and the sampling frequency is synchronized with the above monitoring data; for example, it can be set to monitor every 30 minutes or 60 minutes.

[0019] Simultaneously, satellite surface chlorophyll a image sequences are acquired using remote sensing satellite imaging technology. Medium-resolution satellite imaging platforms, such as MODIS or Sentinel-2 remote sensing satellites, can be used to acquire spatial distribution images of surface chlorophyll a across the entire target reservoir area at various imaging times. Each satellite surface chlorophyll a image represents the chlorophyll a concentration value in pixel form, with units such as milligrams per cubic meter (mg / m³). The images acquired at each imaging time constitute a satellite surface chlorophyll a image sequence, and the imaging time information is recorded for each image.

[0020] A unified time grid is used for time registration. First, a unified time grid is determined; for example, the hourly intervals of each hour are used as unified time grid nodes. Multi-depth water temperature data, multi-depth dissolved oxygen data, multi-depth chlorophyll a data, and inflow / outflow flow data obtained at each sampling time are mapped to the determined unified time grid. Time registration methods, such as linear interpolation or cubic spline interpolation algorithms, are used to match each dataset to the corresponding unified time grid node. For satellite surface chlorophyll a image sequences, interpolation or nearest neighbor matching algorithms are also used to align the image sequences corresponding to each imaging time with the unified time grid.

[0021] After time registration is completed, unified data identification processing is performed. First, time identifiers are established, that is, each node of the unified time grid is assigned a unique time identifier to uniquely indicate the sampling time or imaging time corresponding to the data. Depth identifiers are established, that is, a unique depth identifier is assigned to each monitoring depth. For example, depths of 0.5 meters, 2 meters, 5 meters, 10 meters, 20 meters, and 30 meters below the water surface are assigned fixed depth identifiers. Spatial location identifiers are established. The inflow and outflow flow data are identified by the spatial coordinates of the monitoring section (e.g., latitude and longitude or UTM coordinates). Each pixel of the satellite surface chlorophyll a image sequence is identified by a specific spatial location identifier in the form of latitude and longitude or UTM coordinate grid.

[0022] The above three identifiers are individually bound to their corresponding data records to form a multi-source spatiotemporal dataset. Data binding is manifested in that each data record contains a unified time identifier, depth identifier, and spatial location identifier, corresponding to multi-depth water temperature data, multi-depth dissolved oxygen data, multi-depth chlorophyll a data, inflow and outflow data, and the corresponding chlorophyll a data in the satellite surface chlorophyll a image sequence. For example, each data record can be represented as {time identifier, depth identifier, spatial location identifier, multi-depth water temperature data (unit: degrees Celsius, ℃), multi-depth dissolved oxygen data (unit: milligrams per liter, mg / L), multi-depth chlorophyll a data (unit: milligrams per cubic meter, mg / m³), inflow and outflow data (unit: cubic meters per second, m³ / s), satellite surface chlorophyll a data (unit: milligrams per cubic meter, mg / m³)}.

[0023] S2: Perform Bayesian variable point detection on multi-depth water temperature and dissolved oxygen data to identify thermocline boundaries and oxygen-consuming boundaries, and generate a water column stratification boundary table, including: Based on a unified time identifier, multi-depth water temperature data corresponding to the same time identifier is extracted and sorted according to the unified depth identifier to form a water temperature profile corresponding to the time identifier. Similarly, based on a unified time identifier, multi-depth dissolved oxygen data corresponding to the same time identifier is extracted and sorted according to the unified depth identifier to form a dissolved oxygen profile corresponding to the time identifier. Each water temperature profile is constructed with water temperature as the ordinate (unit, for example, degrees Celsius, ℃) and the depth value corresponding to the depth identifier as the abscissa (unit, for example, meters, m). Each dissolved oxygen profile is constructed with dissolved oxygen concentration as the ordinate (unit, for example, milligrams per liter, mg / L) and the depth value corresponding to the depth identifier as the abscissa (unit, for example, meters, m).

[0024] For the water temperature profile, a Bayesian change point detection method is used to determine the locations of abrupt changes in the water temperature gradient as the boundaries of the thermocline. Specifically, a Bayesian probability model is constructed using multi-depth water temperature data sequences from the water temperature profile as input data to estimate the posterior probability distribution of the water temperature profile. A Markov Chain Monte Carlo (MCMC) algorithm is used to solve the model, with an initial number of iterations (e.g., 10,000). During iteration, the difference between adjacent iteration results is calculated in real time. When the difference between consecutive iterations is less than a set convergence threshold (e.g., 0.001), the iteration stops and the model is considered converged. The posterior probability value of the water temperature gradient in the depth direction is extracted from the model's calculation results. When the posterior probability value at a certain depth is higher than a set change point determination threshold (e.g., 0.9), the corresponding depth location is determined as the thermocline boundary. The posterior probability threshold is set by statistically analyzing historical multi-depth water temperature data to calculate the historical probability distribution of the water temperature gradient, and selecting a probability threshold that can distinguish between abrupt gradient changes and gradual gradient changes using probability density distribution analysis.

[0025] For dissolved oxygen profiles, a Bayesian change point detection method is used to determine the locations of abrupt changes in dissolved oxygen concentration as oxygen consumption boundaries. Specifically, a Bayesian probability model is established to estimate the posterior probability distribution of the dissolved oxygen profile using multi-depth dissolved oxygen data sequences as input data. A Markov Chain Monte Carlo (MCMC) algorithm is used to solve the model, with an initial iteration count set to, for example, 10,000. The same convergence criterion as for the water temperature profile is used: iteration stops when the difference between consecutive iterations is less than a set convergence threshold (e.g., 0.001). The posterior probability value of dissolved oxygen concentration in the depth direction is obtained from the model's calculation results. When the posterior probability value at a certain depth is higher than a set change point determination threshold (e.g., 0.9), the corresponding depth location is determined as the oxygen consumption boundary. The posterior probability threshold for abrupt changes in dissolved oxygen concentration is set as follows: based on statistical analysis of historical multi-depth dissolved oxygen data, the historical probability distribution of dissolved oxygen concentration changes is calculated, and probability density distribution analysis is used to select a probability threshold that can distinguish between abrupt and gradual concentration changes.

[0026] To ensure a temporal correspondence between the determined thermocline and oxygen-consuming boundaries, a water column stratification boundary table is created by associating the thermocline and oxygen-consuming boundaries corresponding to each time marker based on a unified time identifier. Each record in the water column stratification boundary table includes a unified time identifier, a thermocline boundary depth identifier, and an oxygen-consuming boundary depth identifier. Both the thermocline and oxygen-consuming boundary depth identifiers are depth location values ​​(in meters, for example) determined by a Bayesian variable point detection method. For example, it can be represented as: {Time identifier: March 1, 2026, 10:00 AM, Thermocline boundary: 5 meters, Oxygen-consuming boundary: 10 meters}.

[0027] S3: Based on the water column stratification boundary table, multi-depth chlorophyll a data is randomly rearranged within the time blocks corresponding to the same water column stratification boundary to generate a stratified random substitution sequence, including: Based on the water column stratification boundary table, this study analyzes the combined relationships between the thermocline boundary depth identifiers and the oxygen-consuming boundary depth identifiers recorded in the table, using a unified time identifier, to determine the time blocks corresponding to similar water column stratification boundaries. Each record in the water column stratification boundary table is represented as a two-dimensional feature vector, where each depth identifier is in meters (m). The Euclidean distance between the two-dimensional feature vectors of all records is calculated to measure the similarity between different records.

[0028] Based on statistical analysis of historical data, such as calculating the probability density distribution of Euclidean distance between all records, and selecting the quantile of the probability density distribution (e.g., the 75th percentile) as the distance threshold, when the Euclidean distance between two records is less than the threshold, the two records are determined to belong to the same type of water column stratification boundary. Based on the set of time identifiers corresponding to the records determined to be of the same type, the time blocks corresponding to each water column stratification boundary are determined. For example, a time block can be a continuous set of time identifiers between 10:00 AM on March 1, 2026 and 10:00 AM on March 3, 2026, or it can be multiple discontinuous sets of time identifiers distributed at intervals.

[0029] After determining the time blocks corresponding to the stratification boundaries of similar water columns, multi-depth chlorophyll a data corresponding to each time marker within each time block is extracted based on a unified time identifier, constructing a multi-depth chlorophyll a data set within each time block. The data record corresponding to each time marker is accurately located based on the unified time identifier; multi-depth chlorophyll a data corresponding to each depth marker is extracted from each data record, where the unit of multi-depth chlorophyll a data is, for example, milligrams per cubic meter (mg / m³); the extracted data are sorted according to the unified depth identifier, forming a chlorophyll a sequence for each depth with the depth value of the depth identifier as the x-axis (unit: meters, m) and the multi-depth chlorophyll a data as the y-axis (unit: milligrams per cubic meter, mg / m³), ensuring that the chlorophyll a data corresponding to all time markers within each time block forms a structurally consistent sequence set.

[0030] After generating chlorophyll a sequences at each depth, the order of time markers within the corresponding time blocks of each depth's chlorophyll a sequence is randomly rearranged to generate the corresponding hierarchical random substitution sequences. The number of time markers within each time block corresponding to the chlorophyll a sequence at each depth is determined, for example, a time block containing 48 time markers. The random rearrangement process maintains the following constraints: the total number of data points within each time block remains constant; the depth marker of each data point remains fixed; and the time block affiliation of each data point remains unchanged. A random rearrangement algorithm is used to implement the above operations, such as the Fisher-Yates random shuffle algorithm. Starting from the last time marker within the time block, the last time marker is swapped with any previously randomly positioned time marker, and this process is repeated until all time markers are randomly arranged. After the random rearrangement is complete, the data points of each depth's chlorophyll a sequence are rearranged according to the new random time marker order, thus generating the hierarchical random substitution sequences. Each stratified random substitution sequence is completely identical to the corresponding original chlorophyll a sequence at each depth, meaning that the unit of each data point is milligrams per cubic meter (mg / m³), and the corresponding depth identifier is meters (m).

[0031] The number of random rearrangements is determined based on statistical analysis of historical multi-depth chlorophyll a data. For example, the range of rearrangements required to achieve statistical stability of the stratified random substitution sequence can be determined by calculating statistical stability indices (such as changes in standard deviation or mean) in historical random rearrangement experiments. For instance, the number of random rearrangements can be set to 100. Each random rearrangement operation is performed independently and initialized with a different random seed to ensure independence and randomness between each rearrangement process. For example, the random seed can be the current timestamp of the computer system, thereby ensuring the reliability and repeatability of the generated random substitution sequence.

[0032] S4: Calculate the lag window transfer entropy from multi-depth chlorophyll-a data and hierarchical random substitution sequences to satellite surface chlorophyll-a image sequences, and compare them to generate a depth surface information flow table, including: Based on unified time and spatial location identifiers, surface chlorophyll a sequences corresponding to chlorophyll a sequences at various depths are extracted from satellite surface chlorophyll a image sequences. The extraction method is as follows: Based on the unified time identifier, the images in the satellite surface chlorophyll a image sequence corresponding to each time identifier within each time block are determined; based on the spatial location identifier, the pixels corresponding to the spatial location of the monitoring point are located, and the surface chlorophyll a concentration corresponding to each time identifier is extracted, in units such as milligrams per cubic meter (mg / m³); the extracted surface chlorophyll a concentrations are arranged according to the unified time identifier order to form a surface chlorophyll a sequence with the time identifier as the x-axis and the surface chlorophyll a concentration as the y-axis, ensuring accurate temporal matching with the corresponding chlorophyll a sequences at various depths.

[0033] Within the same time block, the lag window transfer entropy between chlorophyll-a sequences at each depth and between the stratified random substitution sequence and the surface chlorophyll-a sequence is calculated to quantitatively characterize the information flow intensity between chlorophyll-a sequences at each depth and the surface chlorophyll-a sequence. The input data required for calculating the lag window transfer entropy are chlorophyll-a sequences at each depth, the stratified random substitution sequence, and the surface chlorophyll-a sequence. The method for determining the lag window length is as follows: mutual information analysis or correlation analysis is performed on historical data, and the lag time corresponding to the occurrence of the maximum correlation coefficient is selected as the window length. Typical lag window lengths are, for example, 1 to 24 hours (h). The lag window transfer entropy is calculated as follows: ; in, The lag window transfer entropy represents the information flow intensity from sequence X to sequence Y, i.e., the information flow intensity from each depth chlorophyll a sequence or hierarchical random substitution sequence to the surface chlorophyll a sequence, in bits. This indicates the surface chlorophyll a data corresponding to the identifier at the next time step; This represents the historical data of surface chlorophyll a sequences corresponding to k consecutive time markers tracing back from the current time marker t; This represents the historical data of chlorophyll-a sequences or stratified random substitution sequences at each depth, traced back l consecutive time markers from the current time marker t; l represents the length of the historical data of chlorophyll-a sequences or stratified random substitution sequences at each depth in the conditional probability used to calculate the transfer entropy, i.e., the number of time markers of the continuous depth chlorophyll-a data considered when tracing back from the current time marker t, for example, set to 1 to 3 time markers; joint probability distribution function. The calculation method is as follows: Define a sliding time window, and use the frequency ratio of data occurrences within the sliding time window as the probability value. The window length is, for example, determined to be 48 time markers to satisfy the stability of probability estimation.

[0034] After calculation, two sets of hysteresis window transfer entropies are obtained from the chlorophyll a sequence at each depth and from the stratified random substitution sequence to the surface chlorophyll a sequence. The unit of the information flow intensity for both sets is bits. For example, for a chlorophyll a sequence with a depth of 5 meters, two transfer entropies can be obtained. One corresponds to the information flow intensity from the actual chlorophyll a sequence at each depth to the surface chlorophyll a sequence, and the other corresponds to the information flow intensity from the stratified random substitution sequence to the surface chlorophyll a sequence.

[0035] The transfer entropy of two sets of lag windows was compared and analyzed according to a unified depth identifier to determine the actual information flow intensity of chlorophyll a sequences relative to random sequences at each depth. The comparison process was as follows: a Monte Carlo permutation significance test was performed on the transfer entropy of the real data and the transfer entropy of the random data under each depth identifier. The random permutation was repeated and the transfer entropy was calculated, for example, repeated 1000 times, to establish the probability distribution of the random transfer entropy. The transfer entropy value of the real sequence was compared with the probability distribution of the transfer entropy of the random sequence. When the transfer entropy value of the real sequence was higher than the 95th percentile of the probability distribution of the transfer entropy of the random sequence, that is, when the significance level was 0.05, the transfer entropy value of the real sequence was determined to be a significantly effective information flow intensity; otherwise, the information flow intensity was determined to be insignificant and marked as zero. The significantly effective lag window transfer entropy was used as the actual information flow intensity corresponding to the chlorophyll a sequence at each depth, in bits.

[0036] The deep surface information flow table is a structured data table containing a uniform time identifier, a depth identifier, and a corresponding information flow intensity, where the unit of information flow intensity is a bit. For example, a typical record in the deep surface information flow table can be represented as: {Time identifier: March 1, 2026, 10:00, Depth identifier: 5 meters, Information flow intensity: 0.35 bits}.

[0037] S5: When the information flow intensity of the surface depth information flow table is lower than a preset threshold, calculate the offset of the chlorophyll a peak depth relative to the thermocline boundary, and estimate the upwelling time of the water layer based on the inflow and outflow data to form an ex-situ deep-layer signal table, including: Based on the deep surface information flow table, a method is determined to filter information flow intensities below a preset threshold. The preset threshold is determined by statistically analyzing the historical information flow intensities in the deep surface information flow table, using probability density function analysis to calculate the quantiles of the information flow intensity, for example, setting the 25th percentile as the preset threshold. For each record, when the corresponding information flow intensity is lower than the determined preset threshold, the corresponding unified time identifier and depth identifier are selected and recorded.

[0038] Based on a defined unified time marker, a multi-depth chlorophyll-a data set corresponding to the unified time marker is extracted to determine the peak depth of chlorophyll-a at the corresponding time. Multi-depth chlorophyll-a data corresponding to all depth markers recorded based on the unified time marker are located, with each data point in milligrams per cubic meter (mg / m³). The extracted multi-depth chlorophyll-a data is sorted sequentially from the surface to the deepest layer according to the unified depth marker, forming a chlorophyll-a depth profile. The chlorophyll-a depth profile uses the depth value corresponding to the depth marker as the x-axis (in meters, m) and the chlorophyll-a concentration value as the y-axis (in milligrams per cubic meter, mg / m³). A sliding window method is used for peak search: a window of a certain length (e.g., a sliding window length of 1 meter) is gradually slid from the surface to the deepest layer, and the maximum chlorophyll-a concentration and its corresponding depth position are recorded during each sliding process. After traversing all data, the depth position containing the global maximum chlorophyll-a concentration is selected from the set of recorded maximum values ​​as the peak depth of chlorophyll-a.

[0039] After determining the peak depth of chlorophyll a, the thermocline boundary depth identifier, identical to the unified time identifier, is extracted from the water column stratification boundary table. The corresponding record in the water column stratification boundary table is then searched based on the unified time identifier, and the depth identifier and corresponding depth value (in meters, m) of the included thermocline boundary are extracted from the record. The offset of the peak chlorophyll a depth relative to the thermocline boundary is calculated by subtracting the thermocline boundary depth from the peak chlorophyll a depth. The sign of the offset is as follows: a positive offset indicates that the peak chlorophyll a is below the thermocline boundary, while a negative offset indicates that it is above the thermocline boundary.

[0040] To accurately estimate the time to water upwelling, inflow and outflow data, expressed in cubic meters per second (m³ / s), were extracted based on a unified time identifier. The extraction method included: matching inflow and outflow data records from a multi-source spatiotemporal dataset with the unified time identifier, and extracting the outflow rate corresponding to the time identifier. The time to water upwelling was calculated using a dynamic model of vertical water migration. The dynamic model formula is as follows: ;in, This indicates the estimated time it will take for the water layer to rise to the surface, in hours (h). This represents the offset of the calculated chlorophyll a peak depth relative to the thermocline boundary, in meters (m). The effective cross-sectional area of ​​the migrating water body (in square meters, m²) is determined based on the cross-sectional geometry and capacity curve of the reservoir at the monitoring point. For example, a typical cross-sectional area is 5000 square meters (m²). This represents the extracted outflow rate, expressed in cubic meters per second (m³ / s). The above formula allows us to calculate the theoretically shortest time for the water layer to migrate upwards from the peak chlorophyll a depth to the thermocline boundary.

[0041] If offset When the value is negative, indicating that the chlorophyll a peak is already above the thermocline boundary, the water layer uplift time is set to 0 hours (h), meaning no migration time calculation is performed. If the extracted outflow is 0 or abnormally low (e.g., less than the flow value at the 5th percentile of historical flow data), the calculated water layer uplift time is set to the maximum allowable time (e.g., 24 hours), indicating that effective vertical migration cannot be achieved under the current flow conditions. An ex-situ deep signal table is constructed, where each record includes a unified time identifier, a depth identifier, the offset of the chlorophyll a peak depth relative to the thermocline boundary (in meters, m), and the corresponding calculated water layer uplift time (in hours, h).

[0042] S6: Construct positive and negative samples based on the deep surface information flow table and the ectopic deep signal table, train a deep learning network, and output a surface chlorophyll a prediction map and water quality warning level, including: Based on the deep surface information flow table, a unified time identifier and depth identifier corresponding to an information flow intensity not lower than a preset threshold are determined to construct a positive sample data set for deep learning network training. The preset threshold for information flow intensity is set as follows: The information flow intensity distribution of historical data is calculated from the historical information flow intensity data in the deep surface information flow table, and the statistical quantiles of the information flow intensity are determined through probability density function statistical analysis. For example, the 75th percentile of historical information flow intensity is selected as the preset threshold. For each record in the deep surface information flow table, when the information flow intensity of the record is greater than or equal to the preset threshold, the corresponding unified time identifier and depth identifier are extracted and recorded as positive sample identifier information.

[0043] Based on the determined positive sample identifier information, multi-depth chlorophyll-a data and surface chlorophyll-a sequences corresponding to the unified time identifier of the positive samples are extracted from the multi-source spatiotemporal dataset to form the positive sample dataset. The extraction of multi-depth chlorophyll-a data specifically involves: locating all corresponding depth identifier positions in the multi-source spatiotemporal dataset according to the determined unified time identifier, and extracting multi-depth chlorophyll-a data corresponding to each depth identifier. The extracted data unit is milligrams per cubic meter (mg / m³). After extraction, the data are arranged in the order of the unified depth identifiers to form a structurally consistent multi-depth chlorophyll-a data sequence. Each data record is represented as {unified time identifier, multi-depth chlorophyll-a data}. The method for extracting the surface chlorophyll a sequence is as follows: Based on the same unified time identifier and spatial location identifier, the corresponding spatial location pixels of the satellite surface chlorophyll a image sequence are located, and the corresponding chlorophyll a concentration data is extracted from them, with the unit being milligrams per cubic meter (mg / m³). The data are arranged in the order of the unified time identifier to form a complete surface chlorophyll a sequence data, and each data record is {unified time identifier, surface chlorophyll a sequence}.

[0044] Based on the ex-situ deep signal table, a method for constructing a negative sample dataset is proposed: Negative sample identification information is extracted from the ex-situ deep signal table, including a unified time identifier, depth identifier, the offset of the chlorophyll a peak depth relative to the thermocline boundary, and the corresponding water upwelling reachable time. Then, based on the negative sample identification information, multi-depth chlorophyll a data with corresponding time identifiers is extracted from a multi-source spatiotemporal dataset. This information is then combined with the offset and water upwelling reachable time to form a negative sample data structure. The negative sample data structure is: {Unified time identifier, depth identifier, multi-depth chlorophyll a data (unit: milligrams per cubic meter, mg / m³), offset (unit: meters, m), water upwelling reachable time (unit: hours, h)}.

[0045] The constructed positive and negative sample datasets are used for training the deep learning network. The network selection method is as follows: based on the spatial and temporal characteristics of this water quality prediction task, a suitable deep learning network for fusing time-series information and spatial distribution characteristics is selected, such as a Long Short-Term Memory (LSTM) network or a Convolutional LSM network. The network structure includes an input layer, hidden layers, an output layer, and corresponding mismatch constraint structures. The network input consists of multi-depth chlorophyll-a data and surface chlorophyll-a sequences from the positive sample dataset, and multi-depth chlorophyll-a data and corresponding offsets and water layer uplift reach times from the negative sample dataset.

[0046] The training method for the deep learning network is as follows: positive sample data is used as the positive prediction input of the network, with the surface chlorophyll a sequence as the supervision label and multi-depth chlorophyll a data as the network input features; negative sample data is used as the mismatch constraint input to constrain network overfitting. This is achieved by defining a loss function that includes mismatch constraints, where the loss function is the mean square error between the network's predicted output and the actual surface chlorophyll a sequence, and adding a negative sample mismatch constraint penalty term, which is the error term between the network's predicted negative sample output and the offset and the water layer upturn reach time; the network training parameters are as follows: the initial learning rate is set to, for example, 0.001, the Adam algorithm is selected as the optimization algorithm, the number of network training iterations is, for example, 1000 times, and the stopping condition is: the change in the loss function value is less than 0.0001 after 50 consecutive iterations.

[0047] The trained deep learning network is used for actual surface chlorophyll a prediction. Based on the multi-source spatiotemporal dataset corresponding to the time period to be predicted, the input is fed into the network for prediction. The network outputs a spatial prediction map of surface chlorophyll a, represented in a two-dimensional spatial grid format. Each grid cell represents the predicted chlorophyll a concentration value at the corresponding spatial location, in milligrams per cubic meter (mg / m³). The spatial resolution of the prediction map is, for example, 10 meters × 10 meters.

[0048] The water quality warning level is generated based on the spatial prediction map of surface chlorophyll a. Specifically, based on the correspondence between historical chlorophyll a concentration and water quality risk level, a concentration range is set. For example, 0 to 2 mg / m³ is set as low risk level, 2 to 5 mg / m³ as medium risk level, and greater than 5 mg / m³ as high risk level. The predicted concentration value of each spatial grid is mapped to the risk level according to the concentration range to form the final spatial distribution map of water quality warning level. Example 2

[0049] The difference between Embodiment 2 and Embodiment 1 is that this embodiment introduces a water quality prediction system based on deep fusion and deep learning.

[0050] Figure 2 A schematic diagram of the water quality prediction system based on deep fusion and deep learning of the present invention is given. The water quality prediction system based on deep fusion and deep learning includes: Data acquisition module: acquires multi-depth water temperature data, multi-depth dissolved oxygen data, multi-depth chlorophyll a data, inflow and outflow flow data, and satellite surface chlorophyll a image sequence, and generates multi-source spatiotemporal datasets according to time grids; Layer identification module: Performs Bayesian variable point detection on multi-depth water temperature data and multi-depth dissolved oxygen data, identifies thermocline boundaries and oxygen consumption boundaries, and generates a water column layer boundary table; Sequence rearrangement module: Based on the water column stratification boundary table, it performs block random rearrangement on multi-depth chlorophyll a data within the time blocks corresponding to the water column stratification boundaries of the same type, generating a stratified random substitution sequence; Information flow table module: Calculates the lag window transfer entropy from multi-depth chlorophyll a data and hierarchical random substitution sequence to satellite surface chlorophyll a image sequence, and compares and generates a depth surface information flow table; Anomaly detection module: When the information flow intensity of the surface depth information flow table is lower than the preset threshold, calculate the offset of the chlorophyll a peak depth relative to the thermocline boundary, and estimate the upwelling time of the water layer based on the inflow and outflow flow data to form an anomaly deep layer signal table. Prediction and early warning module: Construct positive and negative samples based on the deep surface information flow table and the ectopic deep signal table, train a deep learning network, and output a surface chlorophyll a prediction map and water quality early warning level.

[0051] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0052] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0053] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0054] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0055] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0056] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0057] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0058] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0059] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A water quality prediction method based on deep fusion and deep learning, characterized in that, Includes the following steps: S1: Acquire multi-depth water temperature data, multi-depth dissolved oxygen data, multi-depth chlorophyll a data, inflow and outflow data, and satellite surface chlorophyll a image sequence, and generate a multi-source spatiotemporal dataset according to the time grid; S2: Perform Bayesian variable point detection on multi-depth water temperature data and multi-depth dissolved oxygen data to identify thermocline boundaries and oxygen consumption boundaries, and generate a water column stratification boundary table; S3: Based on the water column stratification boundary table, the multi-depth chlorophyll a data is randomly rearranged within the time blocks corresponding to the water column stratification boundaries of the same type to generate a stratified random substitution sequence; the thermocline boundary depth identifier and oxygen consumption boundary depth identifier of each record in the water column stratification boundary table are represented as two-dimensional feature vectors, and the Euclidean distance between the two-dimensional feature vectors of all records is calculated. When the Euclidean distance between two records is less than the judgment distance threshold, the two records are judged to belong to the water column stratification boundary of the same type. S4: Calculate the lag window transfer entropy from multi-depth chlorophyll-a data and stratified random substitution sequences to satellite surface chlorophyll-a image sequences, and compare them to generate a depth surface information flow table; under the same depth identifier, compare the lag window transfer entropy from each depth chlorophyll-a sequence to the surface chlorophyll-a sequence with the lag window transfer entropy from the corresponding stratified random substitution sequence to the surface chlorophyll-a sequence; the depth surface information flow table is a structured data table containing a unified time identifier, depth identifier, and corresponding information flow intensity; S5: When the information flow intensity of the surface depth information flow table is lower than a preset threshold, calculate the offset of the chlorophyll a peak depth relative to the thermocline boundary, and estimate the upwelling time of the water layer based on the inflow and outflow data to form an ectopic deep-layer signal table; the upwelling time of the water layer is calculated using a dynamic model of vertical water migration, and the dynamic model formula is: ;in, This indicates the estimated time it will take for the water layer to rise to the surface. This represents the offset of the calculated chlorophyll a peak depth relative to the thermocline boundary; Indicates the effective cross-sectional area of ​​the migrating water body; This indicates the extracted outflow rate; the ex-situ deep signal table, in which each record includes a unified time identifier, a depth identifier, the offset of the chlorophyll a peak depth relative to the thermocline boundary, and the corresponding calculated water layer upturn reach time; S6: Construct positive and negative samples based on the deep surface information flow table and the ectopic deep signal table, train a deep learning network, and output a surface chlorophyll a prediction map and water quality warning level, specifically: Based on the information flow table of depth and surface, filter the time identifier and depth identifier corresponding to the information flow intensity not lower than the preset threshold, and extract the corresponding multi-depth chlorophyll a data and surface chlorophyll a sequence to construct positive samples. Based on the ectopic deep signal table, the corresponding time marker, depth marker, offset, and water upturn reachable time are selected to extract the corresponding multi-depth chlorophyll a data, and negative samples are constructed by combining the offset and water upturn reachable time. The deep learning network is trained by using positive samples as surface prediction inputs and negative samples as mismatch constraint inputs. The multi-source spatiotemporal dataset is input into the trained deep learning network, which outputs a prediction map of surface chlorophyll a and a water quality warning level.

2. The water quality prediction method based on deep fusion and deep learning according to claim 1, characterized in that, S1, specifically: Collect multi-depth water temperature data, multi-depth dissolved oxygen data, and multi-depth chlorophyll a data at different monitoring depths and at each sampling time; Acquire inflow and outflow data and satellite surface chlorophyll a image sequences corresponding to each sampling time; Time registration was performed on each sampling and imaging time according to a unified time grid, and time markers, depth markers, and spatial location markers were established for multi-depth water temperature data, multi-depth dissolved oxygen data, multi-depth chlorophyll a data, inflow and outflow flow data, and satellite surface chlorophyll a image sequences to form a multi-source spatiotemporal dataset.

3. The water quality prediction method based on deep fusion and deep learning according to claim 2, characterized in that, S2, specifically: Based on the multi-depth water temperature data and multi-depth dissolved oxygen data corresponding to the same time marker, water temperature profiles and dissolved oxygen profiles are formed according to the depth markers; Bayesian variable point detection was performed on the water temperature profile along the depth direction to determine the locations of abrupt changes in the water temperature gradient as the boundaries of the thermocline. Bayesian variable point detection was performed on the dissolved oxygen profile along the depth direction to determine the locations of abrupt changes in dissolved oxygen concentration as oxygen consumption boundaries. A water column stratification boundary table is generated by associating the thermocline boundary and oxygen consumption boundary with the time marker.

4. The water quality prediction method based on deep fusion and deep learning according to claim 3, characterized in that, S3, specifically: Based on the water column stratification boundary table, the time blocks corresponding to the same type of water column stratification boundary are determined according to the correspondence between the thermocline boundary and the oxygen consumption boundary. Extract multi-depth chlorophyll a data corresponding to time markers within the time block, and form chlorophyll a sequences for each depth according to the depth markers; Within the time blocks corresponding to chlorophyll a sequences at each depth, the number of data, depth identifiers, and time block affiliations remain unchanged. The order of the time identifiers is randomly rearranged to generate a hierarchical random substitution sequence.

5. The water quality prediction method based on deep fusion and deep learning according to claim 4, characterized in that, S4, specifically: Based on time and spatial location identifiers, surface chlorophyll a sequences corresponding to chlorophyll a sequences at each depth are extracted from satellite surface chlorophyll a image sequences. Within the same time block, calculate the hysteresis window transfer entropy from each depth chlorophyll a sequence and the stratified random substitution sequence to the surface chlorophyll a sequence; By comparing the lag window transfer entropy with the depth identifier, the information flow intensity corresponding to the chlorophyll a sequence at each depth is determined, and a depth surface information flow table is generated.

6. The water quality prediction method based on deep fusion and deep learning according to claim 5, characterized in that, S5, specifically: Based on the deep surface information flow table, filter the time identifier and depth identifier corresponding to information flow intensity below a preset threshold; Multi-depth chlorophyll a data is extracted based on time markers to determine the peak depth of chlorophyll a. The thermocline boundary is extracted by combining the water column stratification boundary table, and the offset of the peak depth of chlorophyll a relative to the thermocline boundary is calculated. Based on the time marker associated with the inflow and outflow data, the time when the water layer can reach the top is estimated by combining the offset of the chlorophyll a peak depth relative to the thermocline boundary. An ex-situ deep-layer signal table is formed based on time markers, depth markers, offsets, and the time when the water layer can be upturned.

7. A water quality prediction system based on deep fusion and deep learning, used to implement the water quality prediction method based on deep fusion and deep learning as described in any one of claims 1-6, characterized in that, include: Data acquisition module: acquires multi-depth water temperature data, multi-depth dissolved oxygen data, multi-depth chlorophyll a data, inflow and outflow flow data, and satellite surface chlorophyll a image sequence, and generates multi-source spatiotemporal datasets according to time grids; Layer identification module: Performs Bayesian variable point detection on multi-depth water temperature data and multi-depth dissolved oxygen data, identifies thermocline boundaries and oxygen consumption boundaries, and generates a water column layer boundary table; Sequence rearrangement module: Based on the water column stratification boundary table, it performs block random rearrangement on multi-depth chlorophyll a data within the time blocks corresponding to the water column stratification boundaries of the same type, generating a stratified random substitution sequence; Information flow table module: Calculates the lag window transfer entropy from multi-depth chlorophyll a data and hierarchical random substitution sequence to satellite surface chlorophyll a image sequence, and compares and generates a depth surface information flow table; Anomaly detection module: When the information flow intensity of the surface depth information flow table is lower than the preset threshold, calculate the offset of the chlorophyll a peak depth relative to the thermocline boundary, and estimate the upwelling time of the water layer based on the inflow and outflow flow data to form an anomaly deep layer signal table. Prediction and early warning module: Construct positive and negative samples based on the deep surface information flow table and the ectopic deep signal table, train a deep learning network, and output a surface chlorophyll a prediction map and water quality early warning level.

Images