Marine environmental protection-oriented multi-source heterogeneous data retrieval method and system

Abstract

The invention relates to the technical field of distributed storage information retrieval, in particular to a multi-source heterogeneous data retrieval method and system for marine environmental protection, and the method comprises the steps: constructing a marine data ontology library, and determining a cross-modal feature vector mapping space through a double-flow neural network; space-time reference unification is executed, and a standardized retrieval unit is packaged; enhancing the optical image based on the underwater physical model and generating a multi-modal joint index key; constructing a hybrid index structure to manage the three-dimensional spatio-temporal data; utilizing Bloom filter fingerprints and binary exclusive-OR logic to execute sea-land increment index synchronization; and generating an extended query vector based on the marine knowledge graph, and outputting a result in combination with cross-modal similarity matching and space-time constraint. According to the method, the technical defects of marine environment data modal isomerism, semantic association deficiency, non-uniform three-dimensional distribution and synchronous lag in a narrow bandwidth environment are overcome by constructing a deep fusion distributed retrieval architecture, and the retrieval precision ratio and the sea-land data collaboration efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of distributed storage information retrieval technology, specifically to a method and system for retrieving multi-source heterogeneous data for marine environmental protection. Background Technology

[0002] With the rapid development of marine observation technology, marine environmental protection and monitoring have formed a three-dimensional sensing network encompassing satellite remote sensing, UAV patrols, automatic identification systems for ships, and underwater robots, generating massive, rapidly updated, and diverse "5V big data." To address the management needs of this massive amount of data, existing technologies widely employ distributed storage architectures to store this multi-source, heterogeneous data. In existing technological systems, a divide-and-conquer strategy is typically used: for structured time-series data such as temperature, salinity, and depth, relational databases combined with general spatial indexes are often used for management; for semi-structured scientific data such as NetCDF / HDF5, distributed file systems are typically used for storage and indexing via metadata header information; while unstructured data such as seabed videos are often archived as independent files.

[0003] However, the aforementioned existing technologies have significant shortcomings when facing the unique complexity of the marine environment. Their core technical problem lies in the inability of general distributed storage and indexing mechanisms to adapt to the extreme non-uniformity of marine data distribution in time and space, as well as the semantic fragmentation between multimodal data. Marine data is extremely unevenly distributed in both vertical depth and horizontal geographic space. Directly applying general octrees or R-trees on top of distributed storage leads to a severely unbalanced index structure, wasting storage space and prolonging cross-node query response times. For high-frequency streaming data such as AIS, maintaining real-time updates of traditional indexes in a distributed environment incurs expensive node splitting and consistency synchronization overhead. Due to the lack of a unified semantic mapping mechanism, a semantic gap exists between the low-level visual features of unstructured data in distributed storage and the high-level semantic concepts queried by users, making efficient and accurate joint retrieval between heterogeneous data impossible.

[0004] Therefore, this invention proposes a method and system for retrieving multi-source heterogeneous data for marine environmental protection. Summary of the Invention

[0005] The purpose of this invention is to provide a retrieval method and system for multi-source heterogeneous data for marine environmental protection. By constructing a cross-modal vector projection mechanism based on physical attribute parameter normalization, multi-source heterogeneous features are mapped to a unified semantic space to eliminate semantic fragmentation. Furthermore, an adaptive hybrid spatiotemporal index structure based on water depth gradient is established. The dual-layer architecture of memory and disk is used to adapt to the spatiotemporal non-uniform distribution of marine data, thereby effectively solving the technical problem that the general distributed storage and indexing mechanism proposed in the background technology cannot adapt to the extreme non-uniformity of the spatiotemporal distribution of marine data and the semantic fragmentation between multimodal data.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for retrieving multi-source heterogeneous data for marine environmental protection includes: At edge nodes, temperature and salinity profile data, ship trajectory data, ocean grid data, optical image data, and environmental data are collected to construct an ocean data ontology library and a cross-modal feature vector mapping space, generate standardized retrieval units, and parse transmission constraint parameters from the optical image data. The optical image data is transformed using transmission constraint parameters to generate a visual retrieval code. The visual retrieval code is then associated with the environmental data using a cross-modal feature vector mapping space to generate a multimodal joint index key, which is then written into a standardized retrieval unit. Standardized retrieval units are stored in a distributed manner, and a hybrid index structure is built. Index segments are periodically merged to generate a distributed storage index. Generate a Bloom filter fingerprint for the distributed storage index, receive a reference fingerprint from the shore-based central node, detect fingerprint differences between the distributed storage index and the reference fingerprint, and transmit only incremental index data blocks extracted based on fingerprint differences via a remote communication link.

[0007] Preferably, the construction of the marine data ontology library and cross-modal feature vector mapping space specifically includes: collecting metadata fields from edge nodes such as temperature and salinity profile data, ship trajectory data, marine raster data, and environmental data; establishing a three-layer ontology structure including a concept layer, an attribute layer, and an instance layer; defining the hierarchical relationships and spatial topological relationships between monitored element entities to form a marine data ontology library; collecting optical image data as visual samples and environmental data as text samples to construct a two-stream neural network structure including a visual encoding branch and a text encoding branch; inputting visual samples into the visual encoding branch to output visual feature vectors, and inputting text samples into the text encoding branch to output text feature vectors; calculating the cosine similarity loss value between the visual feature vector and the text feature vector; updating the weight parameters of the two-stream neural network structure through a backpropagation algorithm until the cosine similarity loss value is lower than a preset threshold; and establishing the trained two-stream neural network structure as a cross-modal feature vector mapping space.

[0008] Preferably, the generation of standardized retrieval units specifically includes: mapping the spatiotemporal attribute fields of temperature and salinity profile data, ship trajectory data, marine raster data, optical image data, and environmental data to a unified spatiotemporal reference using a marine data ontology library; converting the timestamps of all data to Coordinated Universal Time (UTC) format and the spatial coordinates to the WGS84 geographic coordinate system; using the Laida criterion to detect and remove numerical outliers in the temperature and salinity profile data and ship trajectory data, and performing linear interpolation to fill in missing time-series data; and encapsulating the cleaned data into standard data units containing a unified header identifier, spatiotemporal coding fields, and data payload, and assembling them to form standardized retrieval units.

[0009] Preferably, the step of generating a multimodal joint index key and writing it into the standardized retrieval unit specifically includes: parsing the water attenuation coefficient and background light intensity from the optical image data as underwater channel transmission parameters; based on the underwater dark channel prior logic, performing red light channel compensation and descattering calculations on the optical image data using the water attenuation coefficient, and outputting a normalized image matrix; inputting the normalized image matrix into the visual encoding branch of the cross-modal feature vector mapping space, and extracting visual feature vectors as visual retrieval codes; inputting environmental data that is spatiotemporally aligned with the optical image data into the text encoding branch of the cross-modal feature vector mapping space, and extracting text feature vectors; performing a concatenation operation between the visual feature vectors and the text feature vectors to generate a multimodal joint index key, and storing the multimodal joint index key in the index field of the standardized retrieval unit.

[0010] Preferably, the construction of the hybrid index structure specifically includes: constructing a memory layer for the hybrid index structure, allocating a MemTable storage area in memory as the first level of the LSM-Tree structure, receiving real-time arriving ship trajectory data and corresponding multimodal composite index keys, and writing the data to the SSTable file on the disk when the MemTable storage area reaches the capacity threshold; constructing a disk layer for the hybrid index structure, constructing an octree structure on the disk as a global spatial partitioning index, dividing the ocean geographic space along the depth axis into multiple water depth levels, and establishing octree nodes for each water depth level; determining the amount of data within the leaf nodes of the octree, and when the amount of data exceeds the splitting threshold, constructing an R-tree structure nested inside the leaf nodes of the octree, and using the R-tree structure to manage the remaining data and multimodal composite index keys within the spatial range of the leaf nodes; maintaining the timeliness of the hybrid index structure, starting a background merging process, periodically reading the SSTable file and the R-tree structure within the leaf nodes of the octree, performing merge sorting based on timestamps and spatial coordinates, deleting redundant version data, and generating a distributed storage index.

[0011] Preferably, the process of generating the Bloom filter fingerprint of the distributed storage index specifically includes: traversing the index data blocks in the distributed storage index and extracting the unique identifier of the index data block; performing hash calculation on the unique identifier using a preset independent hash function to obtain a hash value; creating a binary bit array of fixed length and setting the bits corresponding to multiple hash value positions in the binary bit array to 1; using the processed binary bit array as the Bloom filter fingerprint of the index data block, and aggregating the Bloom filter fingerprints of multiple index data blocks to form a node-level fingerprint set.

[0012] Preferably, the transmission of incremental index data blocks based on fingerprint difference extraction via the remote communication link specifically includes: reading the Bloom filter fingerprint of the distributed storage index at the edge node and receiving a reference fingerprint from the shore-based central node deployed in the terrestrial data center; performing a binary XOR operation on the Bloom filter fingerprint and the reference fingerprint to identify the location index of the bit value difference; locating the newly added index data block in the distributed storage index using a reverse mapping table based on the location index, and marking it as an incremental index data block; performing binary serialization and compression encoding on the incremental index data block, and sending it to the shore-based central node via the remote communication link transmission protocol.

[0013] Preferably, the step of generating extended query vectors using a marine knowledge graph specifically includes: using the marine data ontology as the graph mode layer, traversing the standardized retrieval units, reading the seawater temperature, salinity, and depth profile data and ship dynamic trajectory flow data contained in the standardized retrieval units, and parsing out geographic entity names, monitoring equipment identifiers, and marine organism categories to generate graph instance nodes; calculating the topological inclusion relationship of the graph instance nodes in the spatiotemporal dimension, and extracting the causal evolution relationship between entities from the marine environmental monitoring text contained in the standardized retrieval units, establishing semantic connection edges, and forming a marine knowledge graph; performing word segmentation and named entity recognition on the received retrieval instructions to extract retrieval keywords; locating the target instance node matching the retrieval keywords in the marine knowledge graph, traversing all one-hop neighbor nodes of the target instance node along the semantic connection edges, and filtering out two-hop neighbor nodes with causal evolution relationships to form an associated node set; using the text encoding branch of the cross-modal feature vector mapping space, converting the retrieval keywords and the associated node set into high-dimensional semantic vectors respectively; performing weighted average calculation on the high-dimensional semantic vectors to generate extended query vectors.

[0014] Preferably, the step of matching and outputting multi-source data retrieval results in the distributed storage index specifically includes: calculating the cosine similarity between the extended query vector and the multimodal joint index key stored in the distributed storage index; selecting multimodal joint index keys with a cosine similarity higher than the matching threshold as a candidate key set; parsing the spatiotemporal encoding field corresponding to the candidate key set, and using a hybrid index structure to filter candidate keys that do not conform to the spatiotemporal range of the retrieval instruction; reading the corresponding standardized retrieval unit from the distributed storage according to the filtered candidate keys, and outputting the multi-source data retrieval results in order of cosine similarity value.

[0015] A retrieval system for multi-source heterogeneous data for marine environmental protection includes: The data acquisition and standardization module is configured to collect temperature and salinity profile data, ship trajectory data, ocean grid data, optical image data and environmental data at edge nodes, construct an ocean data ontology library and a cross-modal feature vector mapping space, generate standardized retrieval units, and parse transmission constraint parameters from the optical image data. The multimodal joint index generation module is configured to transform optical image data using transmission constraint parameters to generate visual retrieval codes, and to associate the visual retrieval codes with environmental data using cross-modal feature vector mapping space to generate multimodal joint index keys and write them into standardized retrieval units. The distributed storage and hybrid index building module is configured to perform distributed storage on standardized retrieval units, build a hybrid index structure, and periodically merge index segments to generate a distributed storage index. The edge-center incremental synchronization module is configured to generate a Bloom filter fingerprint for the distributed storage index, receive a reference fingerprint from the shore-based central node, detect fingerprint differences between the distributed storage index and the reference fingerprint, and transmit only incremental index data blocks extracted based on fingerprint differences through the remote communication link. The semantic expansion and joint retrieval module is configured to receive retrieval instructions, generate extended query vectors using the marine knowledge graph, match and output multi-source data retrieval results in the distributed storage index.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention utilizes transmission constraint parameters to transform optical image data to generate visual retrieval codes, and establishes a vector space association between the transformed optical image data and environmental data using a cross-modal feature vector mapping space, generating a multimodal joint index key. This method eliminates numerical interference to visual signals from complex underwater channel environments (such as light attenuation and acoustic scattering) during data preprocessing by utilizing transmission constraint parameters. Furthermore, through a unified vector space mapping association, it establishes an index path between the underlying encoding of unstructured image data and the semantic encoding of structured environmental data, achieving joint retrieval and interoperability of cross-modal data. This effectively solves the technical challenges of semantic association failure and low recall rates caused by signal degradation in specific marine environments using traditional retrieval methods.

[0017] 2. This invention constructs a hybrid index structure and performs distributed storage on standardized retrieval units. This method adaptively adapts to the physical distribution characteristics of marine data through a hybrid index mechanism, balancing the real-time writing requirements of high-frequency streaming data such as ship trajectories with the multi-dimensional query requirements of massive historical archive data. It avoids the node vacancy and tree height imbalance phenomena caused by a single general-purpose index (traditional R-tree) in sparse marine regions, significantly improving the storage space utilization and multi-dimensional retrieval response speed of massive non-uniform spatiotemporal data.

[0018] 3. This invention generates a Bloom filter fingerprint of a distributed storage index, detects its fingerprint differences with the reference fingerprint of the shore-based central node, and transmits only incremental index data blocks based on difference extraction via a remote communication link. This method leverages the high compression ratio of Bloom filters to significantly reduce metadata communication overhead during synchronization, abandons the full index replication mode, and greatly reduces bandwidth consumption on remote communication links such as satellites. While ensuring eventual consistency between edge and central index data, it significantly improves the timeliness and stability of the shore-based center acquiring the latest monitoring data from the edge. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for retrieving multi-source heterogeneous data for marine environmental protection proposed in this invention; Figure 2 This is a schematic diagram of the structure of a multi-source heterogeneous data retrieval system for marine environmental protection proposed in this invention; Figure 3 This is a flowchart of a land-sea incremental index synchronization method based on Bloom filter fingerprint, which is a retrieval method for multi-source heterogeneous data for marine environmental protection proposed in this invention. Detailed Implementation

[0020] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention provides a method and system for retrieving multi-source heterogeneous data for marine environmental protection. The technical solution is as follows: A method for retrieving multi-source heterogeneous data for marine environmental protection, referencing Figure 1 The specific implementation steps of the method proposed in this invention include: At edge nodes, temperature and salinity profile data, ship trajectory data, ocean grid data, optical image data, and environmental data are collected to construct an ocean data ontology library and a cross-modal feature vector mapping space, generate standardized retrieval units, and parse transmission constraint parameters from the optical image data. The optical image data is transformed using transmission constraint parameters to generate a visual retrieval code. The visual retrieval code is then associated with the environmental data using a cross-modal feature vector mapping space to generate a multimodal joint index key, which is then written into a standardized retrieval unit. Standardized retrieval units are stored in a distributed manner, and a hybrid index structure is built. Index segments are periodically merged to generate a distributed storage index. Generate a Bloom filter fingerprint for the distributed storage index, receive a reference fingerprint from the shore-based central node, detect fingerprint differences between the distributed storage index and the reference fingerprint, and transmit only incremental index data blocks extracted based on fingerprint differences via a remote communication link.

[0022] Example 1: This embodiment provides a specific application of a method for retrieving multi-source heterogeneous data for marine environmental protection.

[0023] Furthermore, metadata fields from edge nodes, including temperature and salinity profile data, ship trajectory data, marine raster data, and environmental data, are collected to establish a three-layer ontology structure comprising a conceptual layer, an attribute layer, and an instance layer. The hierarchical relationships and spatial topological relationships between monitored entity entities are defined to form a marine data ontology library. Optical image data is collected as visual samples, and environmental data is collected as text samples to construct a two-stream neural network structure containing visual encoding and text encoding branches. Visual samples are input into the visual encoding branch to output visual feature vectors, and text samples are input into the text encoding branch to output text feature vectors. The cosine similarity loss value between the visual feature vector and the text feature vector is calculated, and the weight parameters of the two-stream neural network structure are updated through backpropagation until the cosine similarity loss value is lower than a preset threshold. The trained two-stream neural network structure is then established as a cross-modal feature vector mapping space.

[0024] Specifically, in the three-layer ontology structure, the concept layer defines four primary core classes: marine hydrology, marine life, marine pollution, and human activities, and defines hierarchical subclasses under each core class (such as temperature, salinity, and pressure subclasses under marine hydrology); the attribute layer defines the spatial logical relationships between entities (such as location, intersection, and containment) as well as descriptive attributes such as data precision, unit, and acquisition frequency.

[0025] In the dual-stream neural network, the visual encoding branch adopts an improved ResNet-50 architecture, replacing the convolutional layer with a stride of 2 with dilated convolutions to expand the receptive field. The input size is fixed at 224×224 pixels, and after passing through a global average pooling layer, it is connected to a 1024-dimensional fully connected mapping layer. The text encoding branch is based on the Transformer encoder structure, setting the maximum sequence length to 128, and mapping the text semantics to the same 1024-dimensional vector space as the visual features.

[0026] The text encoding branch uses a RoBERTa-Large pre-trained model, mapping the classification vectors from the last hidden layer to 1024 dimensions using a non-linear activation function. During training, mean squared error combined with a cosine similarity loss function is used for optimization, with a learning rate set to [value missing]. When the cosine similarity on the validation set is higher than 0.96 and the loss function tends to be stable, training is stopped and all hidden layer weights are frozen. By saving the model weight file (.pth or .bin), a cross-modal feature vector mapping space that can project the visual distribution of images and the semantic distribution of text onto the same high-dimensional hypersphere is formally established.

[0027] The loss function uses a weighted combination of contrastive loss and cosine similarity loss, and the loss value L is calculated as follows: ; in, Let i be the i-th visual feature vector. For the corresponding positive sample (matching text features). This means that within a training batch, excluding positive sample pairs... In addition, with visual vectors The text feature vectors with the highest cosine similarity (i.e., the hard negative samples) are used to force the model to learn more fine-grained semantic distinguishing features. This is the interval threshold (with a value of 0.2). The formula aims to minimize... This reduces the cosine distance between positive sample pairs, while the max function ensures that the similarity of positive sample pairs is at least higher than that of negative sample pairs. .

[0028] A semantic skeleton for marine data was constructed using an ontology library. Combined with a two-stream network, heterogeneous data was mapped to the same high-dimensional quantitative space, solving the problem of inability to directly correlate and compare data caused by the semantic gap between different modalities.

[0029] Furthermore, the spatiotemporal attribute fields of temperature and salinity profile data, ship trajectory data, marine raster data, optical imagery data, and environmental data are mapped to a unified spatiotemporal reference using a marine data ontology library; the timestamps of all data are converted to Coordinated Universal Time (UTC) format, and the spatial coordinates are converted to the WGS84 geographic coordinate system; the Laida criterion is used to detect and remove numerical outliers in the temperature and salinity profile data and ship trajectory data, and linear interpolation is performed to fill in missing time-series data; the cleaned data is encapsulated into standard data units containing a unified header identifier, spatiotemporal coding fields, and data payload, and these are aggregated to form standardized retrieval units.

[0030] Specifically, the spatiotemporal mapping logic is implemented through the coordinate transformation module in the ontology library, converting the sensor's built-in projected coordinate system into WGS84 geodetic coordinates (B, L, H). The specific parameters of the Laida criterion are set as follows: collect data samples within a sliding window W=32, and calculate the real-time sample standard deviation. For satisfying Outliers are physically removed, and the removed positions are filled using a three-point linear interpolation algorithm. The standardized retrieval unit adopts a serialized storage structure, with a 16-byte header identifier (containing the Magic Number and protocol version). The spatiotemporal encoding field is composited using 64-bit Geohash and 32-bit vertical depth encoding, and the data payload is aligned with the ontology structure to ultimately generate a standard data packet with self-describing characteristics.

[0031] This ensures that marine environmental protection data from diverse sources are comparable in terms of time and space, while filtering out environmental noise to provide clean and uniformly formatted data input for subsequent large-scale index construction.

[0032] The system retrieves attribute layer constraint rules from the marine data ontology database, extracts the threshold range of physical variables that match the vertical physical depth code of the current standardized retrieval unit, determines whether the predicted value generated by linear interpolation is within the threshold range, and if the interpolation result exceeds the range, it performs weighted correction based on the observation values ​​of nearby monitoring devices within the same time section.

[0033] In the process of generating standardized retrieval units, outliers are eliminated using the Laida criterion, and physical constraints from the ocean data ontology are introduced. When data is missing at a specific vertical physical depth using a temperature profiler, simple linear interpolation may produce drastic fluctuations that do not conform to scientific common sense. In this case, the system automatically queries the extreme values ​​of temperature changes at that depth level defined in the ontology and refers to real-time monitoring data from other buoys or nodes in the same sea area to dynamically correct the interpolation function.

[0034] The solution directly addresses the sampling noise problem caused by the extremely uneven distribution of marine data at vertical depth. Through ontology constraints, it ensures that data collected in extreme environments has a high degree of physical consistency during the preprocessing stage, thereby solving the technical bottleneck of low precision in traditional retrieval methods mentioned in the background, which is caused by poor underlying data quality.

[0035] Furthermore, the water attenuation coefficient and background light intensity are extracted from the optical image data as underwater channel transmission parameters. Based on the underwater dark channel prior logic, the red light channel compensation and descattering calculations are performed on the optical image data using the water attenuation coefficient, outputting a normalized image matrix. The normalized image matrix is ​​input into the visual encoding branch of the cross-modal feature vector mapping space to extract visual feature vectors as visual retrieval codes. Environmental data that is spatiotemporally aligned with the optical image data is input into the text encoding branch of the cross-modal feature vector mapping space to extract text feature vectors. The visual feature vectors and text feature vectors are concatenated to generate a multimodal joint index key, which is then stored in the index field of the standardized retrieval unit.

[0036] The specific steps of parsing the water body attenuation coefficient and background light intensity from optical image data include: calculating the dark channel map of the optical image, selecting the pixels with the highest brightness values ​​in the dark channel map (top 0.1%), calculating their average intensity in the original image as the background light intensity; using the dark channel map and background light intensity to estimate the rough transmittance, and refining the transmittance through guided filtering, thereby inferring the water body attenuation coefficient.

[0037] Specifically, using the optical attenuation model: ; The underwater channel parameters are analyzed, where L is the medium transmission path length (object distance). The object distance L is obtained by using an unsupervised learning-based underwater depth estimation network. This network uses continuous frame video streams of the same scene as input, trains by minimizing photometric reconstruction error and smoothness loss, outputs a normalized relative depth map, and combines sonar ranging data to correct the relative depth to the absolute physical distance. The vertical depth H is only used for the spatial mounting of the retrieval unit and the normalization of global environmental parameters; alternatively, an object distance map can be generated using a monocular depth estimation model. This is specifically for the red band (…). The high attenuation characteristics of the depth map are used to calculate its compensation component, and the specific compensation formula is as follows: ; Where t is transmittance and A is global ambient light. Specifically, the depth map is obtained by reading the vertical depth encoding field from the temperature-salinity profile data that is spatiotemporally aligned with the optical image data in the standardized retrieval unit. Since the standardized retrieval unit has unified the spatial coordinates to the WGS84 geographic coordinate system and performed 32-bit vertical depth encoding, by finding the depth values ​​under the same timestamp and horizontal Geohash encoding, a physical depth parameter d is assigned to each sampling area in the image matrix, and then the transmittance is calculated. The spatial positioning parameters and physical calculation parameters in this invention are clearly distinguished: the vertical physical depth in the geographic coordinate system is denoted as H, which is used for the spatiotemporal encoding and octree index mounting of the standardized retrieval unit; the propagation path length of light underwater (object distance) in the optical attenuation model is denoted as L. The object distance L is not obtained based on the absolute depth H, but rather by performing pixel-level analysis of the original optical image through a monocular depth estimation model integrated with edge nodes to generate an object distance map reflecting the true distance from the object to the lens. After global contrast stretching, the normalized matrix of the descattered image matrix is ​​input into the visual encoding branch of the aforementioned two-stream network to extract a 1024-dimensional visual feature vector. Simultaneously, matching environmental monitoring metadata (such as text descriptions of turbidity, oxygen content, etc.) is extracted and sent to the text encoding branch to obtain a 1024-dimensional text feature vector. The two vectors are concatenated row by row to form a 2048-bit feature fingerprint, which is stored as a multimodal joint index key in the index extension area of ​​the standardized retrieval unit.

[0038] Image quality enhancement was performed for complex marine water conditions, significantly improving the feature recognition rate in low visibility environments. Multimodal feature fusion gives the index strong anti-interference and uniqueness.

[0039] By utilizing the text encoding branch of the cross-modal feature vector mapping space, spatiotemporally aligned salinity and turbidity features are extracted from the payload of the standardized retrieval unit; the attenuation coefficient value in the optical attenuation model is dynamically adjusted according to the salinity and turbidity features; and the compensation component is calculated by combining the object distance information parsed from the image matrix to achieve adaptive enhancement of the red light component of the image.

[0040] This solution enables the collaborative calculation of object distance and environmental parameters. It extracts environmental monitoring text that is aligned with the current image capture time and vertical physical depth from standardized retrieval units. Using the text branch in a dual-stream neural network, these environmental features are converted into correction coefficients, and the attenuation parameters of the red band are dynamically adjusted. When calculating the compensation matrix, it simultaneously refers to the object distance information in the micro-dimensional dimension and the water quality environmental information in the macro-dimensional dimension. The generated normalized image matrix can more realistically restore colors.

[0041] By providing real-time feedback on the physical compensation model using environmental parameters, the feature stability of visual retrieval coding under different turbidity sea areas is significantly improved, effectively addressing the semantic gap between unstructured image data and structured environmental data caused by environmental interference.

[0042] Furthermore, a memory layer with a hybrid index structure is constructed. A MemTable storage area is allocated in memory as the first level of the LSM-Tree structure to receive real-time ship trajectory data and corresponding multimodal composite index keys. When the MemTable storage area reaches its capacity threshold, the data is flushed to the SSTable file on disk. A disk layer with a hybrid index structure is also constructed. An octree structure is built on disk as a global spatial partitioning index, dividing the ocean geographic space along the depth axis into multiple water depth levels. Octree nodes are established for each water depth level. The amount of data within the leaf nodes of the octree is determined. When the amount of data exceeds the splitting threshold, an R-tree structure is nested inside the leaf nodes of the octree. The R-tree structure is used to manage the remaining data and multimodal composite index keys within the spatial range of the leaf nodes. The timeliness of the hybrid index structure is maintained by starting a background merging process, periodically reading the SSTable file and the R-tree structure within the leaf nodes of the octree, performing merge sorting based on timestamps and spatial coordinates, deleting redundant versions of data, and generating a distributed storage index.

[0043] The mounting logic of octree and SSTable is implemented as follows: Each leaf node of the octree not only corresponds to a three-dimensional voxel space, but also stores the start and end values ​​of the 3D-Z-order encoding corresponding to the space range.

[0044] Since the data within SSTable is stored in 3D-Z-order, the leaf nodes of the octree achieve spatial mapping by storing offsets pointing to specific data blocks in the SSTable file. When retrieving a specific sea area, the relevant leaf nodes are first determined through the octree, and then the relevant ordered key-value range within the SSTable is directly located, avoiding a full table scan.

[0045] Specifically, the MemTable layer employs an efficient multi-level ordered skip list implementation, with a memory threshold set at 256MB. SSTable files written to disk are stored using block compression and mounted using an octree as the space manager. The octree performs non-uniform partitioning for ocean depths (0-11000m), with an initial split threshold of 10000 records. When the data volume within the 3D ocean area corresponding to a leaf node of the octree becomes overloaded, an R+ tree index is constructed within that node, precisely wrapping complex ship trajectory segments using the minimum bounding rectangle. The background merging process uses a time-first strategy, comparing different versions of the same data key. If a record's timestamp is older and exceeds its lifespan, it is physically removed from the SSTable and R-tree, thus establishing a dynamically updated distributed storage index.

[0046] The hybrid index structure balances the high-concurrency writing efficiency of massive streaming trajectory data with the response speed of complex three-dimensional sea area queries, effectively solving the index skew problem caused by uneven distribution of marine vertical depth data.

[0047] Furthermore, the index data blocks in the distributed storage index are traversed to extract the unique identifier of the index data block; the unique identifier is hashed using a preset independent hash function to obtain the hash value; a fixed-length binary bit array is created, and the bits corresponding to multiple hash value positions in the binary bit array are set to 1; the processed binary bit array is used as the Bloom filter fingerprint of the index data block, and the Bloom filter fingerprints of multiple index data blocks are aggregated to form a node-level fingerprint set.

[0048] Specifically, all SSTable file blocks are traversed using an iterator, and a unique identifier for each block is extracted as its identifier. Three independent hash functions (MurmurHash3, CityHash128, and SpookyHash) are pre-defined to map each unique identifier to a 16384-bit bit array. Whenever a data block is indexed, the corresponding three hash slots in the array are set to 1. The aggregation process uses bitwise OR logic to compress thousands of block-level bitmaps on edge nodes into a single node-level fingerprint set. This digest-style representation allows for a compression ratio of over 1:10000 to characterize the data storage status of a node.

[0049] By using a Bloom filter, massive index entries are compressed to the extreme, eliminating the need to load the original index blocks when performing a retrieval preview, which greatly reduces disk I / O and memory pressure.

[0050] Furthermore, the edge node reads the Bloom filter fingerprint of the distributed storage index and receives the reference fingerprint from the shore-based central node deployed in the terrestrial data center; performs a binary XOR operation on the Bloom filter fingerprint and the reference fingerprint to identify the location index of the bit value difference; based on the location index, uses a reverse mapping table to locate the newly added index data block in the distributed storage index and marks it as an incremental index data block; performs binary serialization and compression encoding on the incremental index data block and sends it to the shore-based central node through the remote communication link transmission protocol.

[0051] Specifically, edge nodes periodically synchronize a 2KB node-level fingerprint bitmap to the central node. This is achieved through calculation: ; Locate all bits in the Result that are 1. Use the inverse mapping table in memory (mapping relationship: BitIndex->[BlockID_1, BlockID_2,...]); The system quickly identifies the set of missing data blocks at the central endpoint. A temporary list of index blocks is maintained at the edge nodes. When a Bloom filter detects fingerprint differences, the specific incremental data blocks requiring synchronization are determined by comparing the block digests in the list. These sets are marked as incremental blocks to be synchronized, and the data stream is compressed using a lossless data compression algorithm. The compressed stream is then encapsulated and sent to the land-based central endpoint via a message queue telemetry transmission protocol or satellite communication protocol (such as satellite communication optimized for marine environments) that supports breakpoint resumption, completing the physical closed-loop synchronization of the index. A binary XOR operation is performed to identify the edge fingerprint. With reference fingerprint The difference bitmap Result is analyzed, the index of the bit with a value of 1 in Result is parsed, and the index is then verified twice to be on the edge fingerprint. The original value in the file. If this bit is in The middle is 1 and in If the value in the middle is 0, the difference is determined to be caused by newly added data at the edge node. In this case, the newly added index data block is precisely located and extracted using the in-memory reverse mapping table (BitIndex->BlockID) for serialization and transmission. This logic eliminates interference from central-end-specific data in the synchronization process, ensuring high efficiency of unidirectional incremental transmission in narrow bandwidth environments. The location using the reverse mapping table specifically includes: during the generation of the Bloom filter fingerprint at the edge node, a temporary hash mapping table is maintained in memory. This table uses the output bit index of the hash function as the key and the unique identifier of the index data block as the value. After determining the difference bit index through XOR operation, the corresponding unique identifier is searched in the temporary hash mapping table to locate the newly added index data block.

[0052] It solved the communication problem of extremely limited and unstable bandwidth in ocean long-distance navigation operations, reduced synchronization traffic by more than 90%, and ensured data consistency between shore-based centers and edge observation points.

[0053] Furthermore, using the marine data ontology as the graph mode layer, the standardized retrieval units are traversed, and the seawater temperature, salinity, and depth profile data and ship dynamic trajectory flow data contained in the standardized retrieval units are read. Geographic entity names, monitoring equipment identifiers, and marine organism categories are parsed from these data to generate graph instance nodes. The topological inclusion relationship of the graph instance nodes in the spatiotemporal dimension is calculated, and the causal evolution relationship between entities is extracted from the marine environmental monitoring text contained in the standardized retrieval units to establish semantic connection edges, forming a marine knowledge graph. The received retrieval instructions are segmented and named entity recognition is performed to extract retrieval keywords. The target instance node matching the retrieval keywords is located in the marine knowledge graph, and all one-hop neighbor nodes of the target instance node are traversed along the semantic connection edges. Two-hop neighbor nodes with causal evolution relationships are selected to form an associated node set. Using the text encoding branch of the cross-modal feature vector mapping space, the retrieval keywords and the associated node set are converted into high-dimensional semantic vectors respectively. The high-dimensional semantic vectors are weighted and averaged to generate an extended query vector.

[0054] Specifically, the knowledge graph is constructed based on the schema definition of the ontology library. A Bi-LSTM+CRF model is used to automatically extract four core entities—location, equipment, species, and pollutants—from standard retrieval units. Spatial edges between entities are calculated using the Geohash neighborhood algorithm, while logical edges are determined by text sentiment semantic analysis to determine causal strength. Dependency parsing is used to extract verb-object phrase pairs from the text, and combined with a pre-defined marine environmental causal logic lexicon (causing, inducing, driving), the co-occurrence frequency and time-lag correlation between entities within the same time series are calculated to determine the weight values ​​of causal evolution edges.

[0055] When the search query is "red tide evolution in a certain area of ​​the South China Sea", the system locates the keyword "red tide" and searches for related neighboring nodes (such as "sudden drop in dissolved oxygen" and "algal bloom") along the semantic edges of the evolution in the graph. The query vector is then expanded. The calculation logic is as follows: ; Where K is the search keyword, =0.6. For k-hop neighbor nodes Its weight The weight of a one-hop neighbor decreases exponentially with the number of hops, i.e., the weight of a one-hop neighbor decreases. =0.3, weight of two-hop causal neighbors =0.1. When filtering two-hop nodes, only paths with causal evolutionary semantic connections to one-hop nodes are traversed to filter out irrelevant topological associations.

[0056] It has achieved a leap from keyword matching to semantic reasoning retrieval, enabling retrieval results to cover the causes and consequences of environmental evolution and significantly improving the recall rate of marine environmental protection emergency response.

[0057] The weighting factor of the causal evolution relationship between entities is extracted from the environmental monitoring text contained in the standardized retrieval unit; the logical connectivity strength between the retrieval keyword and the two-hop neighbor node in the time series is calculated; if the strength is lower than the preset drift suppression threshold, the corresponding associated node is removed in the process of generating the extended query vector.

[0058] To prevent deviations in retrieval intent when generating extended query vectors using knowledge graphs, this scheme introduces strong constraints based on causal evolution logic. When the search term "oil spill" reveals a two-hop related node of fish mortality, the timestamps of the two events in the standardized retrieval unit are automatically compared. If the causal evolution weight defined in the graph is low, or if the observation records show that there is no reasonable evolutionary step size between the two in space and time, the interfering node will be removed, and only the monitored element entities with strong logical connections will be retained for the final high-dimensional semantic space matching.

[0059] This solution addresses the drawback of excessive but poorly relevant recall results that often occur in large-scale distributed retrieval. By precisely filtering based on causal evolution strength, it ensures that the retrieval results not only reproduce the evolutionary path of environmental events but also maintain extremely high semantic clustering consistency, significantly improving the retrieval precision in marine environmental emergency response scenarios.

[0060] Furthermore, the cosine similarity between the extended query vector and the multimodal composite index keys stored in the distributed storage index is calculated; multimodal composite index keys with cosine similarity higher than the matching threshold are selected as candidate key sets; the spatiotemporal encoding fields corresponding to the candidate key sets are parsed, and candidate keys that do not conform to the spatiotemporal range of the retrieval instruction are filtered using a hybrid index structure; the corresponding standardized retrieval units are read from the distributed storage according to the filtered candidate keys, and the multi-source data retrieval results are output in order of cosine similarity value.

[0061] Specifically, the Faiss high-dimensional vector index library is invoked to perform a global similarity search within the distributed index, and the expanded query vector is calculated. The cosine distance to the 2048-dimensional joint key in the index field is calculated. The initial matching threshold is set to 0.82. For the generated candidate set, its spatiotemporal encoding is parsed, and Boolean intersection is performed with the spatiotemporal polygon specified in the retrieval command. The fast filtering capabilities of octrees and R-trees are used to remove false matches outside the spatial range. The results are ranked Top-K based on similarity scores, and the standardized retrieval unit parsing module is called to extract the original optical images, trajectory maps, and temperature and salinity data, which are then output to the end user in a structured and visualized manner.

[0062] The process of filtering candidate keys that do not conform to the spatiotemporal range of the retrieval instruction using a hybrid index structure specifically includes: firstly, calculating the encoding of the octree nodes involved based on the spatiotemporal range of the retrieval instruction; for octree nodes that are completely contained within the retrieval range, directly reading all the data blocks attached to them; for octree leaf nodes that intersect with the boundary of the retrieval range, further traversing their internal R-tree structure, and using the minimum bounding rectangle of the R-tree to determine whether the specific data entry is located within the retrieval polygon; simultaneously, searching the skip list structure in the memory MemTable and merging the real-time data results.

[0063] By combining high-dimensional semantic reasoning with precise filtering in underlying physical spatiotemporal dimensions, the system ensures a high degree of relevance of search results in marine professional logic, greatly improving the precision of heterogeneous data retrieval.

[0064] This invention achieves deep semantic fusion of multi-source heterogeneous data by constructing a three-layer ontology structure and a dual-stream neural network. It significantly improves the quality of underlying data under complex sea conditions by using image enhancement and statistical denoising techniques driven by physical models. Furthermore, it combines a hybrid index structure consisting of LSM-Tree, octree, and R-tree with a Bloom filter fingerprint synchronization mechanism to ensure high timeliness and low overhead of three-dimensional spatial retrieval under high-concurrency writing and narrow-bandwidth communication environments. Finally, through semantic expansion of knowledge graphs and cross-modal feature similarity comparison, it achieves precise intelligent retrieval with causal reasoning capabilities, greatly improving the precision, recall, and efficiency of land-sea collaboration in marine environmental monitoring scenarios.

[0065] Example 2: This embodiment provides a specific application of a multi-source heterogeneous data retrieval system for marine environmental protection.

[0066] Furthermore, the data acquisition and standardization module is configured to collect temperature and salinity profile data, ship trajectory data, ocean grid data, optical image data and environmental data at edge nodes, construct an ocean data ontology library and a cross-modal feature vector mapping space, generate standardized retrieval units, and parse transmission constraint parameters from the optical image data. Specifically, the data acquisition and standardization module integrates various marine sensor driver interfaces and an ontology modeling engine. It pre-installs ontology mapping rules based on a second-generation Web ontology language, used to transform heterogeneous fields read from different physical interfaces into semantically consistent concept nodes. The system's built-in cross-modal feature vector mapping space is constructed in parallel by a deep residual network visual branch and a Transformer architecture text branch. The visual branch input resolution is set to 224×224, and the text branch token length is limited to 128. When processing raw sampling points, this module executes a real-time filter based on the Laida criterion, removing samples exceeding 3... Outlier data with biases. The normalized retrieval unit is generated by a Protobuf serializer, which contains a 64-bit Geohash encoded field to provide a binary description of the physical location outside the semantic level, ensuring that the data is fully normalized before entering the storage layer.

[0067] Furthermore, the multimodal joint index generation module is configured to transform optical image data using transmission constraint parameters to generate visual retrieval codes, and to associate the visual retrieval codes with environmental data using cross-modal feature vector mapping space to generate multimodal joint index keys and write them into standardized retrieval units. Specifically, the multimodal joint index generation module includes an underwater imaging physical correction engine, whose transmission constraint parameters mainly include the water attenuation coefficient calculated based on the underwater scattering model. The module executes the red channel compensation function to perform dehazing and color restoration on the optical image, in conjunction with the global background light intensity A. ; The processed normalized image matrix extracts a 1024-dimensional feature vector through the visual branch of the mapping space. Simultaneously, spatiotemporally aligned environmental metadata (such as descriptions of salinity and turbidity) is converted into a 1024-dimensional semantic vector through the text branch. The module then concatenates the two sets of vectors and executes the operation. Normalization generates a 2048-dimensional high-dimensional sparse feature vector as the composite index key. The composite index key is written into the metadata index area of ​​the standardized retrieval unit and serves as the primary key identifier of the distributed KV store.

[0068] Furthermore, the distributed storage and hybrid index building module is configured to perform distributed storage on standardized retrieval units, build a hybrid index structure, and periodically merge index segments to generate a distributed storage index; Specifically, the distributed storage and hybrid indexing module is composed of an LSM-Tree storage engine, a 3D octree partitioner, and an R-tree indexer. The memory layer allocates a 512MB buffer as a MemTable, using a skip list structure to temporarily store standardized retrieval units and index keys received in real-time. When a disk flush threshold is triggered, the data is converted into SSTable segment files. The disk layer uses an octree to divide the ocean's 3D geographic space (latitude, longitude, and depth) into hierarchical voxel nodes. When the data density within a node exceeds 5000 records / voxel, the module automatically constructs an R-tree index within the node. The tree index manages specific trajectories and points using a minimum bounding rectangle. The background merging process employs a Leveled Compaction strategy, scanning the SSTable and R-tree nodes using a multi-way merge algorithm to remove redundant observations within overlapping timestamps, thereby establishing a logically continuous and physically distributed distributed storage index.

[0069] Furthermore, the edge-center incremental synchronization module is configured to generate a Bloom filter fingerprint for the distributed storage index, receive a reference fingerprint from the shore-based central node, detect fingerprint differences between the distributed storage index and the reference fingerprint, and transmit only incremental index data blocks extracted based on fingerprint differences through the remote communication link. Specifically, the edge-center incremental synchronization module includes a fingerprint generator and a logic operation unit. The fingerprint generator uses the MurmurHash3 algorithm to perform five independent hashes on the universally unique identifier of each indexed SSTable block, mapping it to a 16KB binary bit array (Bloom filter). The module receives a reference bitmap from the shore-based center and performs a bitwise XOR operation. Module Analysis The module uses a bit index with a median value of 1, along with a reverse mapping table in memory, to retrieve the corresponding incremental Block ID. For the identified incremental data block, the module calls the LZ4 compression library to compress the data stream and sends the incremental serialized packet to the shore-based center via an adaptive rate-adjusting remote communication protocol (such as a message queue telemetry transmission protocol for sensor networks), thus completing the dynamic alignment of the remote fingerprint. Furthermore, the semantic expansion and joint retrieval module is configured to receive retrieval instructions, generate extended query vectors using the marine knowledge graph, match and output multi-source data retrieval results in the distributed storage index; Specifically, the semantic expansion and joint retrieval module includes a named entity recognition submodule and a vector retrieval server. The named entity recognition submodule uses a BiLSTM-CRF architecture to extract marine entity terms from user search commands and locates corresponding nodes in a marine knowledge graph built based on an ontology. The module performs a two-hop path traversal along the semantic edges of the graph (such as causality and evolution) to filter the set of related nodes. The query vector is then expanded. The calculation follows a weighted fusion logic, with the original instruction weighting 0.6, one-hop related nodes 0.3, and two-hop causal nodes 0.1. Subsequently, the module calls the Faiss vector acceleration library to perform cosine similarity calculation on the composite index key in the distributed storage index, selecting candidate sets with similarity higher than 0.82. It then uses a three-dimensional spatial range filtering operator to exclude invalid spatiotemporal records, finally presenting the sorted and standardized retrieval units to the user.

[0070] This invention's retrieval system integrates marine ontology semantics, underwater physical model compensation, and a dual-stream neural network to construct a high-fidelity standardized perception and feature extraction system. Utilizing a hybrid index structure nested with LSM-Tree, 3D octrees, and R-trees, it effectively addresses the performance bottlenecks of high-concurrency writing and spatiotemporal retrieval caused by the uneven vertical depth distribution of massive marine data. Employing a Bloom filter fingerprint difference detection mechanism significantly reduces the land-sea synchronous payload under narrow-bandwidth satellite links, ensuring distributed consistency. Combining causal reasoning and semantic expansion with marine knowledge graphs, it achieves a leap from keyword matching to domain logic reasoning, significantly improving the precision, recall, and collaborative efficiency of retrieval for complex heterogeneous data in marine environmental protection decision-making scenarios.

[0071] Example 3: This embodiment details a hybrid index construction and maintenance process for massive dynamic marine data. The core of this process is to solve the retrieval bottleneck caused by the extremely uneven distribution of marine environmental monitoring data in vertical depth through the synergy of LSM-Tree architecture and nested three-dimensional spatial index.

[0072] During the data access phase, edge nodes receive dynamic AIS trajectory streams from the ship and temperature, salinity, and depth (TST) profile data collected by sensors. This real-time data first enters the MemTable area in memory. To ensure write throughput under high concurrency scenarios, MemTable uses a multi-level skip list structure to sort the data. The primary key for sorting is composed of 3D-Z-order encoding, timestamp, and device ID.

[0073] The 3D-Z-order encoding generation logic is as follows: The binary bits corresponding to longitude and latitude are interleaved with the discretized vertical depth bits, thus mapping the three-dimensional spatiotemporal coordinates to one-dimensional numerical values, ensuring the local continuity of three-dimensional space within the one-dimensional ordered structure of the LSM-Tree. When the amount of data in memory reaches a preset threshold of 256MB, an asynchronous disk write operation is triggered, converting the memory data into an immutable disk file, SSTable.

[0074] At the disk storage layer, a three-dimensional spatial octree was constructed as the global skeleton. The octree divides the selected three-dimensional ocean area into eight initial subspaces (voxels). Since the density of ocean data acquisition decreases with increasing depth, the octree employs a non-uniform depth constraint strategy during partitioning: in the surface ocean area (0-200 meters), the octree allows partitioning to the maximum depth (8 layers) to match dense ship activity; while in the deep ocean area below 3000 meters, the octree maintains a larger voxel granularity. Specifically, the non-uniform depth constraint strategy is as follows: in the surface ocean area (0-200 meters), the maximum partitioning level of the octree is set to 8 layers to accommodate dense ship trajectory recordings; in the intermediate-deep ocean area (200-3000 meters), the maximum partitioning level is set to 5 layers; and in the deep ocean area below 3000 meters, the maximum partitioning level is limited to 3 layers. If the number of data records in any leaf node exceeds the split threshold (2000 records) but has not reached the corresponding depth level limit, the split continues; if the level limit has been reached but the data volume is still excessive, the process of building an R-tree index nested inside that node is triggered.

[0075] When the number of standardized retrieval units contained in a leaf node of an octree exceeds the splitting threshold of 2000, the system constructs an R-tree index nested within the 3D rectangular space corresponding to that leaf node. The R-tree uses the minimum bounding rectangle algorithm to geometrically wrap each ship trajectory segment or profile line. This nested structure ensures that global retrieval can be achieved quickly using the octree, while precise local retrieval is achieved through spatial overlap calculations within the leaf nodes using the R-tree.

[0076] To maintain the timeliness of the index, the background merge process periodically re-compiles the SSTable files with R... The data blocks in the tree are sorted using a multi-way merge sort. During this process, the system uses the VersionID in the data unit to compare duplicate observations under the same spatial coordinates, physically deletes redundant versions or expired data marked with tombstones, and finally generates a compact distributed storage index by updating the index offset, which significantly optimizes the response efficiency of three-dimensional spatial range queries.

[0077] Example 4: This embodiment details the semantic extension mechanism used in this invention to improve the accuracy of cross-modal retrieval by utilizing a marine knowledge graph. This mechanism aims to compensate for the semantic deficiencies of single-keyword retrieval by leveraging domain prior knowledge. First, the system uses a marine data ontology database as the schema layer of the graph. It employs a BiLSTM-CRF named entity recognition model to parse instance nodes such as geographic entities (South China Sea basin), monitoring equipment (buoy-A01), and marine organisms (phytoplankton) from the text descriptions of standardized retrieval units. Topological relationships between entities are derived through spatial geometry calculations, such as determining whether the coordinates of a monitoring point are located within a geometric polygon of a specific sea area, thereby establishing topological edges indicating whether the point belongs to or is located within that polygon. Furthermore, the system uses dependency parsing to extract causal evolutionary relationships between entities from environmental monitoring reports. For example, from the text stating that rising sea temperatures lead to massive algal blooms, a positive inductive relationship between rising sea temperatures and algal blooms is identified, and semantic connection edges are established in the graph. The resulting marine knowledge graph possesses the ability to infer environmental evolution paths. Upon receiving the user's search command (oil spill impact assessment), the system first performs word segmentation on the command and accurately locates the core node of the oil spill in the graph. Then, the system initiates a semantic traversal algorithm, first extracting one-hop neighbor nodes directly connected to the oil spill (oil content, surface tension), and further filtering out two-hop neighbor nodes with ecological impact (nearshore fish mortality, mangrove ecosystem damage) through causal relationship edges. To generate the final extended query vector, the system calls the text encoding branch of the cross-modal feature vector mapping space, converting the original keywords and all filtered neighbor node descriptions into 1024-dimensional high-dimensional semantic vectors. Extended Query Vector The synthesis formula is as follows: ; Among them, parameters Set to 0.6 to maintain the subjectivity of the query. Set to 0.4 to introduce relevant context. Weight The weighted average decays exponentially based on the number of hops a node has in the graph. The expanded query vector generated through this weighted average calculation not only includes the user's original keyword features but also incorporates expert logic from the field of marine environmental protection. This enables subsequent cosine similarity matching in the distributed index to retrieve more heterogeneous data results with potential causal relationships.

[0078] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for retrieving multi-source heterogeneous data for marine environmental protection, characterized in that, include: At edge nodes, temperature and salinity profile data, ship trajectory data, ocean grid data, optical image data, and environmental data are collected to construct an ocean data ontology library and a cross-modal feature vector mapping space, generate standardized retrieval units, and parse transmission constraint parameters from the optical image data. The optical image data is transformed using transmission constraint parameters to generate a visual retrieval code. The visual retrieval code is then associated with the environmental data using a cross-modal feature vector mapping space to generate a multimodal joint index key, which is then written into a standardized retrieval unit. Standardized retrieval units are stored in a distributed manner, and a hybrid index structure is built. Index segments are periodically merged to generate a distributed storage index. Generate a Bloom filter fingerprint for the distributed storage index, receive a reference fingerprint from the shore-based central node, detect fingerprint differences between the distributed storage index and the reference fingerprint, and transmit only incremental index data blocks extracted based on fingerprint differences through a remote communication link. Upon receiving a retrieval command, the system generates an extended query vector using a marine knowledge graph, matches the results in a distributed storage index, and outputs the retrieval results from multiple data sources.

2. The method for retrieving multi-source heterogeneous data for marine environmental protection according to claim 1, characterized in that, The construction of the marine data ontology and cross-modal feature vector mapping space specifically includes: collecting metadata fields from edge nodes such as temperature and salinity profile data, ship trajectory data, marine raster data, and environmental data; establishing a three-layer ontology structure including a concept layer, an attribute layer, and an instance layer; defining the hierarchical relationships and spatial topological relationships between monitored element entities to form a marine data ontology; collecting optical image data as visual samples and environmental data as text samples to construct a two-stream neural network structure including a visual encoding branch and a text encoding branch; inputting visual samples into the visual encoding branch to output visual feature vectors, and inputting text samples into the text encoding branch to output text feature vectors; calculating the cosine similarity loss value between the visual feature vector and the text feature vector; updating the weight parameters of the two-stream neural network structure through the backpropagation algorithm until the cosine similarity loss value is lower than a preset threshold; and establishing the trained two-stream neural network structure as a cross-modal feature vector mapping space.

3. The method for retrieving multi-source heterogeneous data for marine environmental protection according to claim 1, characterized in that, The generation of standardized retrieval units specifically includes: mapping the spatiotemporal attribute fields of temperature and salinity profile data, ship trajectory data, marine raster data, optical image data, and environmental data to a unified spatiotemporal reference using a marine data ontology library; converting the timestamps of all data to Coordinated Universal Time (UTC) format and the spatial coordinates to the WGS84 geographic coordinate system; using the Laida criterion to detect and remove numerical outliers in the temperature and salinity profile data and ship trajectory data, and performing linear interpolation to fill in missing time-series data; and encapsulating the cleaned data into standard data units containing a unified header identifier, spatiotemporal coding fields, and data payload, and then assembling them to form standardized retrieval units.

4. The method for retrieving multi-source heterogeneous data for marine environmental protection according to claim 1, characterized in that, The specific steps of generating a multimodal joint index key and writing it into the standardized retrieval unit include: parsing the water attenuation coefficient and background light intensity from the optical image data as underwater channel transmission parameters; based on the underwater dark channel prior logic, performing red light channel compensation and descattering calculations on the optical image data using the water attenuation coefficient to output a normalized image matrix; inputting the normalized image matrix into the visual encoding branch of the cross-modal feature vector mapping space to extract visual feature vectors as visual retrieval codes; inputting environmental data spatiotemporally aligned with the optical image data into the text encoding branch of the cross-modal feature vector mapping space to extract text feature vectors; performing a concatenation operation between the visual feature vectors and the text feature vectors to generate a multimodal joint index key, and storing the multimodal joint index key in the index field of the standardized retrieval unit.

5. The method for retrieving multi-source heterogeneous data for marine environmental protection according to claim 1, characterized in that, The construction of the hybrid index structure specifically includes: constructing a memory layer for the hybrid index structure, allocating a MemTable storage area in memory as the first level of the LSM-Tree structure, receiving real-time arriving ship trajectory data and corresponding multimodal composite index keys, and writing the data to the SSTable file on disk when the MemTable storage area reaches the capacity threshold; constructing a disk layer for the hybrid index structure, building an octree structure on disk as a global spatial partitioning index, dividing the ocean geographic space along the depth axis into multiple water depth levels, and establishing octree nodes for each water depth level; determining the amount of data within the leaf nodes of the octree, and when the amount of data exceeds the splitting threshold, constructing an R-tree structure nested inside the leaf nodes of the octree, and using the R-tree structure to manage the remaining data and multimodal composite index keys within the spatial range of the leaf nodes; maintaining the timeliness of the hybrid index structure, starting a background merging process, periodically reading the SSTable file and the R-tree structure within the leaf nodes of the octree, performing merge sorting based on timestamps and spatial coordinates, deleting redundant version data, and generating a distributed storage index.

6. The method for retrieving multi-source heterogeneous data for marine environmental protection according to claim 1, characterized in that, The specific steps for generating the Bloom filter fingerprint of the distributed storage index include: traversing the index data blocks in the distributed storage index and extracting the unique identifier of the index data block; performing hash calculation on the unique identifier using a preset independent hash function to obtain the hash value; creating a binary bit array of fixed length and setting the bits corresponding to multiple hash value positions in the binary bit array to 1; using the processed binary bit array as the Bloom filter fingerprint of the index data block, and aggregating the Bloom filter fingerprints of multiple index data blocks to form a node-level fingerprint set.

7. The method for retrieving multi-source heterogeneous data for marine environmental protection according to claim 1, characterized in that, The specific steps of transmitting incremental index data blocks based on fingerprint difference extraction via the remote communication link include: reading the Bloom filter fingerprint of the distributed storage index at the edge node and receiving a reference fingerprint from the shore-based central node deployed in the terrestrial data center; performing a binary XOR operation on the Bloom filter fingerprint and the reference fingerprint to identify the location index of the bit value difference; locating the newly added index data block in the distributed storage index using a reverse mapping table based on the location index, and marking it as an incremental index data block; performing binary serialization and compression encoding on the incremental index data block, and sending it to the shore-based central node via the remote communication link transmission protocol.

8. The method for retrieving multi-source heterogeneous data for marine environmental protection according to claim 1, characterized in that, The specific steps for generating extended query vectors using a marine knowledge graph include: using the marine data ontology as the graph mode layer, traversing the standardized retrieval units, reading the seawater temperature, salinity, and depth profile data and ship dynamic trajectory flow data contained in the standardized retrieval units, and parsing out geographic entity names, monitoring equipment identifiers, and marine organism categories to generate graph instance nodes; calculating the topological inclusion relationship of the graph instance nodes in the spatiotemporal dimension, and extracting the causal evolution relationship between entities from the marine environmental monitoring text contained in the standardized retrieval units, establishing semantic connection edges, and forming a marine knowledge graph; performing word segmentation and named entity recognition on the received retrieval instructions to extract retrieval keywords; locating the target instance node matching the retrieval keywords in the marine knowledge graph, traversing all one-hop neighbor nodes of the target instance node along the semantic connection edges, and filtering out two-hop neighbor nodes with causal evolution relationships to form an associated node set; using the text encoding branch of the cross-modal feature vector mapping space to convert the retrieval keywords and the associated node set into high-dimensional semantic vectors respectively; and performing a weighted average calculation on the high-dimensional semantic vectors to generate extended query vectors.

9. The method for retrieving multi-source heterogeneous data for marine environmental protection according to claim 1, characterized in that, The specific steps for matching and outputting multi-source data retrieval results in the distributed storage index include: calculating the cosine similarity between the extended query vector and the multimodal joint index key stored in the distributed storage index; selecting multimodal joint index keys with a cosine similarity higher than the matching threshold as a candidate key set; parsing the spatiotemporal encoding field corresponding to the candidate key set, and using a hybrid index structure to filter candidate keys that do not conform to the spatiotemporal range of the retrieval instruction; reading the corresponding standardized retrieval unit from the distributed storage according to the filtered candidate keys, and outputting the multi-source data retrieval results in order of cosine similarity value.

10. A retrieval system for multi-source heterogeneous data for marine environmental protection, characterized in that, include: The data acquisition and standardization module is configured to collect temperature and salinity profile data, ship trajectory data, ocean grid data, optical image data and environmental data at edge nodes, construct an ocean data ontology library and a cross-modal feature vector mapping space, generate standardized retrieval units, and parse transmission constraint parameters from the optical image data. The multimodal joint index generation module is configured to transform optical image data using transmission constraint parameters to generate visual retrieval codes, and to associate the visual retrieval codes with environmental data using cross-modal feature vector mapping space to generate multimodal joint index keys and write them into standardized retrieval units. The distributed storage and hybrid index building module is configured to perform distributed storage on standardized retrieval units, build a hybrid index structure, and periodically merge index segments to generate a distributed storage index. The edge-center incremental synchronization module is configured to generate a Bloom filter fingerprint for the distributed storage index, receive a reference fingerprint from the shore-based central node, detect fingerprint differences between the distributed storage index and the reference fingerprint, and transmit only incremental index data blocks extracted based on fingerprint differences through the remote communication link. The semantic expansion and joint retrieval module is configured to receive retrieval instructions, generate extended query vectors using the marine knowledge graph, match and output multi-source data retrieval results in the distributed storage index.

Images