Multi-element information acquisition and analysis system based on telemetering drifting buoy

Abstract

The invention discloses a multi-element information acquisition and analysis system based on telemetering drifting buoys, and relates to the technical field of marine environment monitoring. A three-dimensional fluorescence spectrometer, a temperature-salinity-depth sensor, a positioning module and a miniature water sampler are carried on each telemetering drifting buoy; the system is used for collecting algae fluorescence spectrums, water quality parameters and position data in real time, inputting the data into a trained random forest model to output the red tide occurrence probability, triggering multi-level early warning in combination with a preset expert rule engine, linking air and shore-based equipment and a data storage system, and utilizing laboratory true value data obtained by a miniature water sampler to determine the red tide occurrence probability. And performing incremental training on the random forest model every day, thereby dynamically optimizing internal structure parameters of the model. The method effectively improves the prediction probability of the red tide.

Description

Technical Field

[0001] The present invention relates to the technical field of marine environment monitoring, and in particular to a multi-factor information collection and analysis system based on a remote sensing drifting buoy. Background Art

[0002] Red tide disasters pose a serious threat to marine ecology and fishery safety. Traditional red tide prediction methods mainly rely on the limited parameters of fixed single-point telemetry drifting buoys, such as single-point water temperature and salinity sensors, and laboratory manual sampling and analysis, which makes it difficult to achieve dynamic coordination of multiple factors in the marine environment.

[0003] The existing technology has the following significant defects: the traditional system only collects a small amount of physical and chemical indicators such as water temperature and dissolved oxygen, and does not fully utilize the bio-optical characteristics and spatiotemporal motion trajectories, and cannot effectively capture the biological-environmental-spatiotemporal triple driving mechanism of red tide outbreaks; fixed telemetry drifting buoys or shore-based equipment are restricted by geographical location and cannot cover vast sea areas, and cannot track the migration path of red tides, resulting in inaccurate prediction of the probability of red tide outbreaks; the existing technology does not fully explore the spatiotemporal correlation of multi-source data when extracting features, resulting in insufficient distinguishing power of the extracted features.

[0004] Therefore, there is an urgent need for a multi-factor information collection and analysis system based on telemetry drifting buoys. By integrating multimodal payloads to collect bio-optical, physical and chemical, and spatiotemporal data in real time, intelligent algorithms are used to mine the spatiotemporal correlation characteristics of multi-source data, and machine learning models are combined to realize red tide probability prediction, so as to break through the bottleneck of extensive feature extraction and low prediction accuracy of traditional technologies. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a multi-factor information collection and analysis system based on telemetry drifting buoys, which solves the problems of traditional red tide prediction such as single factors, extensive feature extraction and low prediction accuracy.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a multi-factor information collection and analysis system based on a telemetry drifting buoy, comprising: The buoy monitoring module, equipped with a three-dimensional fluorescence spectrometer, a temperature, salinity and depth sensor, a positioning module and a micro water sampler on each telemetry drifting buoy, is used to collect algae fluorescence spectra, water quality parameters and location data in real time; The hybrid prediction module inputs real-time algae fluorescence spectra, water quality parameters, and location data into a trained random forest model to output the probability of red tide occurrence, and combines it with a pre-set expert rule engine to trigger multi-level warnings; The early warning response module links airborne and shore-based equipment and data evidence storage systems based on the triggered multi-level early warnings; The offline calibration module uses laboratory ground truth data obtained from micro water samplers to perform incremental training on the random forest model every day, thereby dynamically optimizing the internal structural parameters of the model.

[0007] As a further solution of the present invention, algae fluorescence spectrum data includes the main components of the fluorescence spectrum, the spectrum similarity score, and the algae density conversion value; the water quality parameter data includes water temperature and change rate, salinity and change rate, dissolved oxygen and change rate; and the location data includes the distance from the estuary or coastline and the moving speed.

[0008] As a further solution of the present invention, the principal components of the fluorescence spectrum are obtained by reducing the dimensionality of the intensity matrix through principal component analysis (PCA), and the spectral similarity score is obtained by calculating the cosine similarity between the real-time spectral vector and the algae spectral vector in the fingerprint library.

[0009] As a further solution of the present invention, when the fluorescence spectrum detects that the algae density is greater than a threshold value, the real-time sampling frequency is increased and the micro water sampler is triggered to collect water samples at the same time.

[0010] As a further solution of the present invention, the specific steps of obtaining a trained random forest model are as follows: Collect algae fluorescence spectra, water quality parameters and location data and convert them into multidimensional vectors; Organize several groups of multidimensional vectors into a feature matrix, where rows represent samples, columns represent features, and labels corresponding to each group of multidimensional vectors, where red tide occurrence is 1 and no red tide occurrence is 0; The Gini index is used to evaluate the gain value of each feature for red tide classification, and the feature that contributes most to reducing the Gini impurity is retained, that is, the feature with a Gini gain greater than the threshold is selected for retention; The Bagging strategy is used to randomly extract 70%-80% of samples from historical data to train a single tree, and some of the filtered features are randomly selected when each tree splits; The trained multiple decision trees are integrated to obtain the trained random forest model.

[0011] As a further solution of the present invention, real-time data is input into the trained random forest model to obtain the predicted probability. If the probability is ≤25%, a blue warning is issued; if 25%<probability≤50%, a yellow warning is issued; if 50%<probability≤75%, an orange warning is issued; if the probability is >75%, a red warning is issued.

[0012] As a further solution of the present invention, the pre-set expert rules specifically include: Rule 1: If the measured algae density is greater than a threshold of 1, the alert level is directly raised to orange, regardless of the model probability. If the alert level is already orange or red, it remains unchanged. Rule 2: If the dissolved oxygen is less than threshold 2 and the salinity is less than threshold 3, the warning level is raised to level 2; Rule 3: If the wind speed is greater than threshold 4 and the model probability is less than 60%, the warning level is reduced by 1 level; Rule 4: If the water temperature is greater than the threshold of 5 and the nutrient salt exceeds the standard by 3 times, the basic probability is multiplied by 1.2; Among them, the threshold 1, threshold 2, threshold 3, threshold 4, and threshold 5 are all set according to the actual marine environment, and the nutrient salt is the sum of the concentrations of total nitrogen TN and total phosphorus TP.

[0013] As a further solution of the present invention, if no less than two expert rules are triggered at the same time, the rule with the highest level is selected for execution according to the rule priority. If the warning level after execution exceeds red, it is fixed to red, and if it is lower than blue, it is fixed to blue. The rule priority is: Rule 1>Rule 2>Rule 3>Rule 4.

[0014] As a further solution of the present invention, if it is a blue warning, the data is pushed to the coastal monitoring station via Beidou short message; If it is a yellow warning, the drone hyperspectral remote sensing will be activated to draw a heat map of the red tide spread; If it is an orange alert, the shore-based modified clay spraying system will be activated; If it is a red alert, a fishing ban order will be automatically triggered and the data will be stored on the blockchain.

[0015] As a further solution of the present invention, the specific process of dynamically optimizing the internal structural parameters of the model is as follows: Real-time sensor data and laboratory data from micro-water samplers are collected daily, and the sensor data at the same sampling moment are matched with the laboratory true value to form a calibration data set; Instead of deleting the trained decision trees, several new trees are added based on historical data and the calibration data of the day. For newly added trees, the feature splitting rules are relearned using the calibration dataset, focusing on correcting areas with large historical errors; For the original tree, the influence of outdated rules is weakened through the weight decay mechanism, reducing its weight in voting; Use 30% of the day's calibration data as the test set and calculate the logarithmic loss of the updated model. If the LogLoss decreases by more than 10% compared to before the update, the parameters are confirmed to be effective. Otherwise, an alarm is triggered, prompting manual intervention and inspection. The updated model parameters will be synchronized to the online prediction random forest model at 0:00 the next day.

[0016] The present invention provides a multi-factor information collection and analysis system based on a telemetry drifting buoy, which has the following beneficial effects compared with the existing technology: (1) The present invention integrates a three-dimensional fluorescence spectrometer, a temperature-salinity-depth sensor, and a Beidou positioning module to collect biological optical characteristics, physical and chemical parameters, and spatiotemporal motion data in real time, constructing a multimodal feature vector. This can effectively capture the biological, environmental, and spatiotemporal driving mechanisms, solving the problem of missed key mechanism judgments caused by the traditional method due to the single element. (2) The present invention adopts a telemetry drifting buoy design that supports autonomous movement with ocean currents and real-time data transmission. It can cover vast sea areas that are difficult to monitor with fixed shore-based, single-point telemetry drifting buoys, and track the migration path of red tides, achieving a leap from static single-point monitoring to dynamic full-area tracking, significantly improving the timeliness of disaster response. (3) The present invention constructs an intelligent feature screening model based on the Gini index, replacing the traditional single feature extraction mode driven by artificial experience, and combines the random forest algorithm to perform nonlinear modeling on multi-source data, effectively solving the problems of insufficient feature discrimination and weak model generalization ability in the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a schematic diagram of the system of the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] like Figure 1 The present invention provides a multi-factor information collection and analysis system based on a telemetry drifting buoy, comprising: The buoy monitoring module, equipped with a three-dimensional fluorescence spectrometer, a temperature, salinity and depth sensor, a positioning module and a micro water sampler on each telemetry drifting buoy, is used to collect algae fluorescence spectra, water quality parameters and location data in real time; Algae fluorescence spectrum data includes fluorescence spectrum principal components, spectrum similarity scores, and algae density conversion values; The specific reasons for obtaining the above features are as follows: The original spectrum is the intensity matrix of the excitation-emission wavelength pair. The key features are extracted by principal component analysis (PCA) dimensionality reduction, which is recorded as {PCAi,i∈[1,m]}, where m represents the number of principal components. The principal component reflects the main variation pattern of the spectrum, such as the main peak intensity of chlorophyll a and the position of the shoulder peak of carotenoids, which can distinguish algae species; The cosine similarity between the real-time spectral vector and the algae spectral vector in the fingerprint library is calculated, and the target algae to which it belongs is determined by comparing the similarity threshold, which is directly related to the toxicity risk. Based on the Lambert-Beer law, fluorescence intensity is converted into cell density through a calibration curve. The density increases rapidly during the exponential growth period, and the outbreak trend can be identified by density values ​​at consecutive time points. Water quality parameter data include water temperature and change rate, salinity and change rate, dissolved oxygen and change rate; The specific reasons for obtaining the above features are as follows: Water temperature affects the activity of algae's photosynthetic enzymes. For example, the optimum temperature for Skeletonema costatum is 20-25°C; growth is restricted beyond this range. A sudden rise in water temperature can trigger water stratification and release nutrients from the bottom sediment. The intensity of the stress can be identified by the rate of change. Low salinity is suitable for the reproduction of Noctiluca and Gymnodinium, while high salinity may trigger diatom blooms. A sudden drop in salinity (such as typhoon rainfall) destroys stratification and promotes the upwelling of nutrients, which can be used as a precursor to a red tide outbreak. High dissolved oxygen indicates vigorous algae growth and is a sign of the development of red tides. Algae die and decompose, consuming oxygen. Low dissolved oxygen indicates that the red tide has entered its decline phase, which may trigger secondary disasters. Position data include distance from the estuary or coastline, and movement speed; The specific reasons for obtaining the above features are as follows: Areas less than 10 km from river mouths are significantly affected by land-based nutrients, and the probability of red tides is higher than in the open ocean. Nearshore telemetry drifting buoy data can predict the likelihood of red tides spreading to aquaculture and tourist areas. For example, the risk of impacting human activities increases significantly when the coastline is less than 5 km away. When the speed is greater than 0.5m / s, the red tide can spread over 40km in 24 hours, and the impact range can be assessed by the speed value; When the fluorescence spectrum detects that the algae density is greater than the threshold, the sampling frequency can be automatically increased from 1 time / hour to 1 time / 5 minutes, and the micro water sampler is triggered to collect 500ml of water sample at the same time. The water sampler has a constant temperature storage function to ensure that the biological activity of the water sample is stable for ≥24 hours; Phytoplankton density is a direct quantitative parameter for red tide formation. When the fluorescence spectrum detects that the algae density exceeds the threshold, it indicates that the phytoplankton has entered an exponential growth period. The time window for a large-scale red tide outbreak is usually 12-48 hours. Increasing the sampling frequency at this time can achieve the following: Capture microscopic dynamic changes, collect spectral data at high frequency, and track algal community succession in real time; Verify data reliability and reduce accidental errors, such as misjudgment caused by short-term sensor interference, through short-term high-frequency sampling to ensure the authenticity of early warning signals; Fluorescence spectroscopy is an in-situ rapid detection technology. Although it can reflect algae density in real time, it cannot provide key information such as algae species classification and toxin content. Therefore, when the density threshold is triggered, the synchronous activation of the micro-water sampler has a dual significance; The spectral data represents the total amount of algae at a single point where the remote sensing drifting buoy is located, and the water samples obtained by the water sampler can be used in the laboratory to verify the algae species composition in the area; The laboratory analysis results of water samples can also be used to calibrate subsequent prediction model parameters and improve the generalization ability of the prediction model.

[0020] The hybrid prediction module inputs real-time algae fluorescence spectra, water quality parameters, and location data into a trained random forest model to output the probability of red tide occurrence, and combines it with a pre-set expert rule engine to trigger multi-level warnings; The specific process of random forest training is: Collect multimodal data, i.e., algal fluorescence spectra, water quality parameters, and location data, and convert them into multidimensional vectors; Organize several groups of multidimensional vectors into a feature matrix, where each row is a sample, each column is a feature, and each group of multidimensional vectors has a label corresponding to it, where red tide occurs as 1 and does not occur as 0; The Gini index was used to evaluate the gain of each feature in red tide classification. The higher the gain, the stronger the discrimination ability. Retain the features that contribute most to reducing the Gini impurity, that is, select the features whose Gini gain is greater than the threshold to retain; During the random forest training process, unimportant features will be naturally filtered out, such as noise dimensions that are not related to red tide, because their Gini gain approaches 0 and they are rarely selected as split nodes by the decision tree; At the same time, the advantages of selecting Gini index dimensionality reduction features also include: (1) it needs to be combined with label information, such as red tide and normal data, so the dimensionality reduction is more targeted; (2) it does not require independent preprocessing steps and is embedded in the random forest training process, making the calculation more efficient; (3) it retains the original features with physical meaning, which is convenient for business interpretation; The Gini index is used to evaluate the discriminative ability of features on the target variable. It should be based on samples that can represent the data distribution. The sample size should be as large as possible, close to the full set or using the full training set to avoid sampling error leading to bias in feature importance calculation. For example, taking the water temperature change rate ΔT as an example, assuming that the optimal splitting threshold is ΔT ≥ 1.5°C / h: Subnode 1, ΔT ≥ 1.5°C / h: 30 red tide + 10 normal (purity is 75%), Gini index: Gini1=1-(0.75^2+0.25^2)=0.375; Subnode 2, ΔT < 1.5°C / h: 20 red tide + 40 normal (purity is 33%), Gini index: Gini2=1-(0.67^2+0.33^2)=0.444; Gini gain = 0.5 - (0.4 × 0.375 + 0.6 × 0.444) = 0.084; The Gini gain threshold is pre-set to 0.07, so the water temperature change rate can be extracted; A bagging strategy is used to randomly extract 70%-80% of samples from historical data to train a single tree. When each tree splits, some of the filtered features, such as 2 or 3 features, are randomly selected to enhance model diversity. Multiple decision trees can be trained according to needs, and the probability of red tide occurrence is output through voting. The bagging strategy of random forest requires random sampling of samples and features to enhance the generalization ability of the model. It allows the number of samples to be less than the full set, and can select 70% of the full set or sample set. Each decision tree in the random forest undergoes two random samplings during training: sample sampling, which randomly extracts a subset of the full sample to train a single tree to reduce variance; and feature sampling, which randomly selects some features to calculate the optimal split point when each node of a single tree splits. The above-mentioned extraction of key features through Gini gain belongs to the dimensionality reduction operation in the preprocessing stage, the purpose of which is to eliminate irrelevant features. The feature sampling in this step is an internal mechanism during model training, the purpose of which is to make each tree focus on different feature combinations and enhance diversity. For example, even if only three features are left after Gini index dimensionality reduction, two of them can still be randomly selected for splitting; The real-time collected screening features are input into the model and the red tide probability is output; Construct a basic mapping of probability and warning level based on historical data quantiles: If the probability is ≤ 25%, which corresponds to the 25th percentile in historical data, there is no red tide or a minor outbreak, and a blue alert is issued; If the probability is 25% < ≤ 50%, it corresponds to the historical medium-to-low risk range, and a small-scale red tide may occur, and a yellow warning will be issued; If the probability is less than 50% and less than or equal to 75%, it corresponds to the historical medium-to-high risk range and the possibility of a medium-scale outbreak, and an orange alert is issued; If the probability is >75%, it corresponds to the historically high risk range, indicating the possibility of a large-scale outbreak or toxic algae dominance, and a red alert is issued; Design an expert rule engine to further modify the random forest output probability, introduce domain knowledge to modify the basic level, and solve the problems of model data dependency bias and missed outlier detection in real time; The specific rule types and triggering conditions are as follows: Rule 1: If the measured algae density exceeds a threshold of 1, the alert is immediately upgraded to orange, regardless of the model probability. If the alert is already orange or red, it remains unchanged. The model may underestimate the risk because the training data does not include the latest algae species. The measured density is direct evidence. Rule 2: If the dissolved oxygen is less than threshold 2 and the salinity is less than threshold 3, the level is raised to level 2. Low dissolved oxygen and a sudden drop in salinity are typical signs of worsening red tides, and the model may not capture short-term drastic changes; Rule 3: If the wind speed is greater than a threshold of 4, indicating sustained strong winds, and the model probability is less than 60%, the level is reduced by 1. Strong winds promote water mixing and reduce the probability of red tide aggregation. The model may overestimate the risk in static data. Rule 4: If the water temperature is greater than the threshold of 5 and the nutrient salt (TN+TP) exceeds the standard by 3 times, the base probability is ×1.2, and high temperature and high nutrient salt are the core driving factors of the red tide outbreak, strengthening the model's response to key conditions; Among them, threshold 1, threshold 2, threshold 3, threshold 4, and threshold 5 are all set by the user according to the actual ocean environment; If at least two of the above rules are triggered simultaneously, the rule with the highest priority will be executed according to the following order: Rule 1 > Rule 2 > Rule 3 > Rule 4; If the corrected level exceeds red, it will be fixed in red; if it is lower than blue, it will be fixed in blue.

[0021] The early warning response module triggers multi-level warnings based on comprehensive prediction probabilities, linking airborne and shore-based equipment and data storage systems; If it is a blue warning, the data will be pushed to the coastal monitoring station via Beidou short message; In marine environments, traditional 4G and 5G signals have a coverage rate of less than 30% 50 kilometers offshore, while Beidou short messages enable global communication without blind spots, making them particularly suitable for sea areas without infrastructure. The blue warning corresponds to a low-risk scenario, and the float is in energy-saving mode. For example, the sampling frequency is only once per hour, and the Beidou module consumes very little power, which can extend the float's flight time. Preliminary abnormal data is encrypted and transmitted via BeiDou, providing traceable original records for subsequent warnings. For example, if a certain indicator rises for three consecutive days, it is easy to verify the initial fluctuation point. If it is a yellow warning, the drone hyperspectral remote sensing will be activated to draw a heat map of the red tide spread; Traditional satellite remote sensing has a resolution of 10-100m and cannot identify red tide patches smaller than 500 square meters. However, the 0.5m resolution of drones can capture micro red tides at the aquaculture cage level, such as algae accumulation around a single cage. By inverting the water temperature gradient through near-infrared spectroscopy and combining it with the spectral characteristics of algae, a three-level thermal stratification of red tide core area, diffusion front, and safe area is generated, effectively improving the accuracy; The drone takes only 90 minutes from takeoff to complete a 100km² scan, 30 times faster than the satellite revisit period (2-3 days), making it ideal for capturing the golden 6-hour spread window during red tide outbreaks. If it is an orange alert, the shore-based modified clay spraying system will be activated; Modified clay aggregates algae through electrical neutralization and netting, making it more environmentally friendly than traditional copper sulfate. The residual clay can serve as a natural component of marine sediments and has low ecotoxicity. Fixed spraying stations have a coverage radius of 15km and can treat 20km² of red tide in a single operation. The cost is less than 50 yuan / ㎡, compared to 120 yuan / ㎡ for unmanned boat spraying, making it suitable for regular emergency response. Since this warning corresponds to the stage when red tide has formed but not yet fully erupted, the surface charge of algae cells is not fully polarized at this time, and the coagulation efficiency of modified clay is the highest; If it is a red alert, a fishing ban will be automatically triggered and the data will be stored on the blockchain; A red alert indicates an outbreak of red tide and excessive toxins. The fishing ban will be directly transmitted to fishing vessels via the BeiDou system, ensuring 100% coverage. The fishing ban order is linked to the real-time data of the float. For example, the algae density is updated every 30 minutes. If it is less than 5×10 4 cells / L, the fishing ban will be lifted in advance to avoid excessive control; Through encryption algorithms and distributed ledgers, blockchain can upload key data that triggers red alerts in real time, such as abnormal wave energy values ​​monitored by floats, red tide area dynamics, and water quality indicators, forming an unalterable record with a unique timestamp and hash value. The offline calibration module uses laboratory-based true data obtained from micro-water samplers, such as true algae density and measured nutrient concentrations, to incrementally train the random forest model daily, thereby dynamically optimizing the model's internal structural parameters. Internal structure parameters include feature splitting threshold, which is the critical value of the feature used to divide the sample at each decision tree node; leaf node output value, which is the predicted value corresponding to the final leaf node of the decision tree; feature importance weight, which is the ranking of the model's dependence on each input feature; The specific process of dynamic optimization of the internal structural parameters of the model is as follows: Collect real-time sensor data and laboratory data from micro-water samplers daily, and match the sensor data at the same sampling moment with the laboratory true value to form a calibration data set, such as 500 labeled samples; Instead of deleting the trained decision trees, several new trees are added based on historical data and the calibration data of the day. The specific update method is as follows: For newly added trees, the feature splitting rules are relearned using the calibration dataset, focusing on correcting areas with large historical errors; for existing trees, the influence of outdated rules is weakened through a weight decay mechanism, such as reducing the voting weight of split nodes that conflict with the calibration data; Use 10% of the day's calibration data as the test set and calculate the logarithmic loss of the updated model. If the LogLoss decreases by more than 10% compared to before the update, the parameters are confirmed to be effective. Otherwise, an alarm is triggered, prompting manual intervention and inspection. At 0:00 the next day, the updated model parameters (such as the splitting threshold and leaf node value of the newly added tree) will be synchronized to the online prediction random forest model.

[0022] Some of the data in the above formulas are dimensionless and numerically calculated. Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0023] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A multi-factor information collection and analysis system based on a telemetry drifting buoy, characterized in that: include: The buoy monitoring module, equipped with a three-dimensional fluorescence spectrometer, a temperature, salinity and depth sensor, a positioning module and a micro water sampler on each telemetry drifting buoy, is used to collect algae fluorescence spectra, water quality parameters and location data in real time; The hybrid prediction module inputs real-time algae fluorescence spectra, water quality parameters, and location data into a trained random forest model to output the probability of red tide occurrence, and combines it with a pre-set expert rule engine to trigger multi-level warnings; The early warning response module links airborne and shore-based equipment and data evidence storage systems based on the triggered multi-level early warnings; The offline calibration module uses laboratory ground truth data obtained from micro water samplers to perform incremental training on the random forest model every day, thereby dynamically optimizing the internal structural parameters of the model.

2. A multi-factor information collection and analysis system based on a remote sensing drifting buoy according to claim 1, characterized in that: Algae fluorescence spectrum data include the main components of fluorescence spectrum, spectrum similarity score, and algae density conversion value. Water quality parameter data include water temperature and change rate, salinity and change rate, dissolved oxygen and change rate. Location data include distance from the estuary or coastline and movement speed.

3. The multi-factor information collection and analysis system based on remote sensing drifting buoy according to claim 1 is characterized in that: The principal components of the fluorescence spectrum are obtained by reducing the dimensionality of the intensity matrix through principal component analysis (PCA), and the spectral similarity score is obtained by calculating the cosine similarity between the real-time spectral vector and the algae spectral vector in the fingerprint library.

4. The multi-factor information collection and analysis system based on remote sensing drifting buoy according to claim 1 is characterized in that: When the fluorescence spectrum detects that the algae density is greater than a threshold, the real-time sampling frequency is increased and the micro water sampler is triggered to collect water samples at the same time.

5. The multi-factor information collection and analysis system based on remote sensing drifting buoy according to claim 1 is characterized in that: The specific steps to obtain the trained random forest model are: Collect algae fluorescence spectra, water quality parameters and location data and convert them into multidimensional vectors; Organize several groups of multidimensional vectors into a feature matrix, where rows represent samples, columns represent features, and labels corresponding to each group of multidimensional vectors, where red tide occurrence is 1 and no red tide occurrence is 0; The Gini index is used to evaluate the gain value of each feature for red tide classification, and the feature that contributes most to reducing the Gini impurity is retained, that is, the feature with a Gini gain greater than the threshold is selected for retention; The Bagging strategy is used to randomly extract 70%-80% of samples from historical data to train a single tree, and some of the filtered features are randomly selected when each tree splits; The trained multiple decision trees are integrated to obtain the trained random forest model.

6. The multi-factor information collection and analysis system based on remote sensing drifting buoy according to claim 1 is characterized in that: Input real-time data into the trained random forest model to obtain the predicted probability. If the probability is ≤25%, a blue warning is issued; if 25%<probability≤50%, a yellow warning is issued; if 50%<probability≤75%, an orange warning is issued; if the probability is >75%, a red warning is issued.

7. The multi-factor information collection and analysis system based on remote sensing drifting buoy according to claim 1 is characterized in that: The pre-set expert rules include: Rule 1: If the measured algae density is greater than a threshold of 1, the alert level is directly raised to orange, regardless of the model probability. If the alert level is already orange or red, it remains unchanged. Rule 2: If the dissolved oxygen is less than threshold 2 and the salinity is less than threshold 3, the warning level is raised to level 2; Rule 3: If the wind speed is greater than threshold 4 and the model probability is less than 60%, the warning level is reduced by 1 level; Rule 4: If the water temperature is greater than the threshold of 5 and the nutrient salt exceeds the standard by 3 times, the basic probability is multiplied by 1.2; Among them, the threshold 1, threshold 2, threshold 3, threshold 4, and threshold 5 are all set according to the actual marine environment, and the nutrient salt is the sum of the concentrations of total nitrogen TN and total phosphorus TP.

8. The multi-factor information collection and analysis system based on remote sensing drifting buoy according to claim 1 is characterized in that: If no less than two expert rules are triggered at the same time, the rule with the highest level will be selected for execution according to the rule priority. If the warning level after execution exceeds red, it will be fixed to red, and if it is lower than blue, it will be fixed to blue. The rule priority is: Rule 1 > Rule 2 > Rule 3 > Rule 4.

9. The multi-factor information collection and analysis system based on remote sensing drifting buoy according to claim 1 is characterized in that: If it is a blue warning, the data will be pushed to the coastal monitoring station via Beidou short message; If it is a yellow warning, the drone hyperspectral remote sensing will be activated to draw a heat map of the red tide spread; If it is an orange alert, the shore-based modified clay spraying system will be activated; If it is a red alert, a fishing ban order will be automatically triggered and the data will be stored on the blockchain.

10. The multi-factor information collection and analysis system based on remote sensing drifting buoy according to claim 1, characterized in that: The specific process of dynamically optimizing the internal structural parameters of the model is as follows: Real-time sensor data and laboratory data from micro-water samplers are collected daily, and the sensor data at the same sampling moment are matched with the laboratory true value to form a calibration data set; Instead of deleting the trained decision trees, several new trees are added based on historical data and the calibration data of the day. For newly added trees, the feature splitting rules are relearned using the calibration dataset, focusing on correcting areas with large historical errors; For the original tree, the influence of outdated rules is weakened through the weight decay mechanism, reducing its weight in voting; Use 30% of the day's calibration data as the test set and calculate the logarithmic loss of the updated model. If the log loss decreases by more than 10% compared to before the update, the parameters are confirmed to be effective. Otherwise, an alarm is triggered, prompting manual intervention and inspection. The updated model parameters will be synchronized to the online prediction random forest model at 0:00 the next day.

Images