Semantic indexing and content retrieval method for remote sensing time series data

Abstract

The present application relates to the field of remote sensing data processing, and discloses a semantic indexing and content retrieval method for remote sensing time series data, comprising the following steps: symbolizing and converting a pixel-level numerical time series to generate a symbol sequence composed of phenomorpheme; obtaining spatial context information of the pixel, and dynamically selecting a matching grammar rule from a phenomorpheme knowledge base according to the information; using the selected rule to analyze the phenomorpheme sequence to generate a time series relationship graph; fusing and integrating all the graph fragments generated by the pixel to construct a global graph index library; and a user performs complex structured semantic retrieval on the graph index library through a structured query request. The present application effectively solves the problem of same-spectrum different things of time series curves by introducing adaptive grammar analysis guided by spatial context, converts low-level numerical sequences into high-level semantic relationship networks, and realizes deep and accurate retrieval of complex processes and events in remote sensing data.

Description

Technical Field

[0001] This invention relates to the field of remote sensing data processing technology, specifically to a semantic indexing and content retrieval method for remote sensing time-series data. Background Technology

[0002] With the development of Earth observation technology, massive amounts of remote sensing time-series imagery data have provided unprecedented opportunities for dynamically monitoring changes on the Earth's surface. How to efficiently and accurately extract valuable information from these dense time series is currently a research hotspot and technical challenge in the field of remote sensing.

[0003] Existing remote sensing time-series data analysis techniques largely rely on morphological matching of numerical curves or classification based on statistical features. For example, methods such as Dynamic Time Warping (DTW) classify or retrieve data by calculating the numerical similarity between time-series curves, but this method is sensitive to noise and computationally expensive. In recent years, deep learning models, represented by Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), have been widely used. They can automatically learn deep features in time-series data and have achieved significant accuracy in tasks such as land use / land cover classification.

[0004] However, these mainstream technologies have inherent limitations in deep semantic understanding and content retrieval. Whether it's traditional similarity metrics or complex deep learning models, their core remains fitting patterns to numerical values ​​themselves. This paradigm struggles to effectively address the challenge of "heterogeneous phenomena," where two morphologically very similar time-series curves may correspond to drastically different land cover types or phenological phenomena in different geographical environments or spatial contexts. Existing methods typically treat all pixel analysis logic equally, lacking an adaptive mechanism to dynamically adjust and optimize interpretation rules based on the specific spatial context of each pixel.

[0005] Furthermore, the output of existing technologies is often discrete classification labels or similarity scores, rather than structured, interpretable semantic information. While deep learning models are powerful, their "black box" nature makes their decision-making process difficult to understand, failing to explicitly express what phenological stages or transformation events occurred in the land cover. This leads to shallow retrieval capabilities, preventing users from formulating complex, structured semantic queries based on entities, events, and relationships, such as "retriev all plots of land that changed from 'woodland' to 'grassland' after a 'fire' event."

[0006] Therefore, there is an urgent need for a new technical solution that can fundamentally transform remote sensing time-series data from low-level numerical representations into high-level, machine-understandable semantic relationship networks, and on this basis, achieve truly structured content retrieval, thereby bridging the gap between massive amounts of remote sensing data and users' deep knowledge needs. Summary of the Invention

[0007] The technical problem to be solved by this invention is that existing remote sensing time series data retrieval methods have problems such as insufficient extraction of semantic information of data content, single retrieval dimension, and difficulty in supporting complex related queries.

[0008] To address the aforementioned technical problems, this invention provides a semantic indexing and content retrieval method for remote sensing time-series data.

[0009] The first aspect of this invention provides a semantic indexing and content retrieval method for remote sensing time-series data, the method comprising the following steps:

[0010] Step S1: Obtain the numerical time series of the target pixels in the remote sensing time series data, denoted as... ,in For pixel identifiers, For time.

[0011] Step S2: Convert the numerical time series Convert to a phenological primitive sequence. This step specifically includes:

[0012] S2.1: Logarithmic Time Series Keypoint detection is performed to identify phenological inflection points that define the boundaries of phenological stages, resulting in an ordered set of keypoint timestamps. .

[0013] S2.2: Use the aforementioned set of key point timestamps numerical time series Divide into one or more ordered phenological segments. , of which paragraph Corresponding time interval Sequence fragments within.

[0014] S2.3: For each phenological section Through encoding function Convert it into a phenological primitive in the form of a quadruple. :

[0015] ;

[0016] in:

[0017] Trend Components From the formula The calculation maps the results to a predefined set of trend symbols.

[0018] curvature components From the formula The calculation maps the result to a predefined set of curvature symbols;

[0019] Relative duration component This refers to the absolute duration of the paragraph. Through quantization function The symbol obtained from the mapping;

[0020] Relative amplitude components It is the absolute amplitude of the paragraph. Through quantization function The symbol obtained from the mapping.

[0021] All phenological primitives are arranged in chronological order, constituting the phenological primitive sequence of the pixel. .

[0022] Step S3: Based on the spatial context corresponding to the pixel, select a matching grammar rule from a preset phenological grammar knowledge base. The spatial context is attribute information extracted from a priori geographic information layer based on the pixel's geographic coordinates. The phenological grammar knowledge base stores multiple sets of grammar rules corresponding to different spatial contexts.

[0023] Step S4: Based on the grammar rules selected in Step S3, the phenological primitive sequence is parsed to generate a temporal relation graph representing the semantic information of the pixel. The grammar rules define the mapping relationship from a specific phenological primitive sequence pattern to the action of generating a specific temporal relation graph. The temporal relation graph consists of entities and relations, where entities are used to describe semantic concepts such as ground features, phenological stages, or events, and relations are used to describe the associations between the entities. The relations may have weights or labels representing the association attributes.

[0024] Step S5: Collect all the temporal relationship graphs generated from the pixels to construct a graph index library. The graph index library is constructed based on graph database technology and includes at least an entity index for indexing entities and an adjacency index for indexing relationships.

[0025] Step S6: Generate a graph query pattern based on the user's input query request, and perform matching in the graph index to obtain search results. This step includes converting the semantic concepts and logical constraints in the query request into query constraints on the entities, relationships, and topological structures in the temporal relationship graph, and executing the query constraints in the graph index.

[0026] The aforementioned scheme converts numerical time series into phenological primitive sequences containing morphological information and introduces a spatial context-driven adaptive grammar for parsing, generating structured temporal relationship graphs. This method enables deep semantic modeling of the content of remote sensing time series data. The same phenological primitive sequence, when associated with different spatial contexts, can be parsed into temporal relationship graphs containing different semantic information by selecting different grammar rules. The final graph index supports complex association queries of entities, relationships, and their attributes, solving the problems of limited retrieval dimensions and superficial semantic understanding in existing technologies.

[0027] A second aspect of the present invention provides a semantic indexing and content retrieval system for remote sensing time-series data, the system comprising:

[0028] The symbolization conversion module is configured to execute steps S1 and S2 in the aforementioned method, obtain the numerical time series of pixels and convert it into a phenological primitive sequence;

[0029] An adaptive grammar parsing module, connected to a symbolization conversion module, is configured to execute steps S3 and S4 in the aforementioned method, select grammar rules based on the spatial context of pixels, and parse the phenological primitive sequence into a temporal relation map.

[0030] The graph index building module is connected to the adaptive grammar parsing module and is configured to execute step S5 in the aforementioned method to build the generated temporal relationship graph into a graph index library.

[0031] The content retrieval module is connected to the graph index construction module and is configured to execute step S6 in the aforementioned method, generate a graph query pattern based on the user query, and perform matching in the graph index library to provide retrieval results.

[0032] This invention provides a semantic indexing and content retrieval method for remote sensing time-series data. It has the following beneficial effects:

[0033] 1. This invention converts numerical time series into phenological primitive sequences containing morphological information and parses them based on preset grammatical rules, thereby achieving a structured and symbolic description of the inherent evolutionary patterns of time series data. This method elevates the object of analysis from discrete numerical points to morphological patterns with inherent structures, overcoming the limitations of traditional methods that can only analyze statistical values ​​and cannot understand process and structural information, and is able to extract deeper semantic information.

[0034] 2. This invention introduces spatial context information of pixels to dynamically select matching grammar rules from a phenological grammar knowledge base. This mechanism enables the same phenological primitive sequence to be parsed into different semantic information under different geographical environments, thereby effectively addressing the technical problems of "different objects with the same spectrum" and "different spectra with the same object" in remote sensing image interpretation, and improving the accuracy and regional adaptability of semantic information extraction.

[0035] 3. This invention defines the output of semantic parsing as a structured temporal relation graph containing entities and relationships, and constructs a corresponding graph index library. This data structure and indexing method enable users to initiate deep association queries on complex logical, temporal, and causal relationships between entities through graph query mode, breaking through the limitation of existing technologies that can only support simple searches based on keywords or single tags, and greatly expanding the dimension and depth of data retrieval. Attached Figure Description

[0036] Figure 1 This is a functional module diagram of a semantic indexing and content retrieval system for remote sensing time-series data according to an embodiment of the present invention;

[0037] Figure 2 This is a flowchart illustrating a semantic indexing and content retrieval method for remote sensing time-series data according to an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram of a spatial adaptive grammar parsing process according to an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of a structured semantic retrieval process according to an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, specific embodiments of the invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0041] See attached document Figure 1 , Figure 1This is a functional block diagram of a semantic indexing and content retrieval system for remote sensing time-series data according to an embodiment of the present invention. The present invention provides a semantic indexing and content retrieval system for remote sensing time-series data, which can be deployed on one or more server computing devices. The computing device includes a processor, a memory, a communication interface, and a bus. The processor, memory, and communication interface communicate with each other via the bus. The processor executes computer program instructions stored in the memory to implement the functions of the method described later.

[0042] The system provided in this embodiment aims to convert pixel-level numerical time-series data into structured semantic information capable of deep relational queries. The system includes:

[0043] The data input and preprocessing module acquires raw remote sensing time-series image data and performs a series of standardized processing steps. These steps include radiometric calibration of the raw images, atmospheric correction to eliminate atmospheric effects, and geometric correction to ensure accurate geographic location. After processing, this module precisely aligns and stacks the multi-temporal images to form a complete and clean numerical time series for each pixel in the temporal dimension. The output of this module forms the data basis for all subsequent processing steps.

[0044] The symbolization conversion module is connected to the data input and preprocessing module. This module receives pixel-level numerical time series data from the data input and preprocessing module. And perform the conversion of it into a phenological primitive sequence. The algorithm involves first detecting key points in the input numerical time series to identify phenological inflection points; then, using these inflection points, the sequence is divided into one or more phenological segments. Finally, for each phenological section... Perform four-dimensional feature encoding to generate a phenological primitive. .in, As a trend component, For curvature components, For relative duration components, This represents the relative amplitude component. The output of this module is a sequence of symbols representing the morphological characteristics of the original numerical sequence.

[0045] The adaptive grammar parsing module is connected to the symbolization conversion module. This module receives the phenological primitive sequence from the symbolization conversion module and obtains the spatial context information corresponding to the current pixel from an external geographic information data source. The core function of this module is to dynamically select a set of matching grammar rules from a pre-defined phenological grammar knowledge base based on spatial context information, and then use the selected rules to parse the input phenological primitive sequence. The output of the parsing is a structured temporal relation graph. It contains entities that describe pixel semantics and their interrelationships.

[0046] The graph indexing module is connected to the adaptive grammar parsing module. This module receives discrete temporal relation graph fragments generated by the adaptive grammar parsing module, targeting each pixel. Its function is to aggregate these fragments and inject them into a global graph database, forming a unified temporal relation graph covering the entire study area. Simultaneously, this module indexes the nodes (entities) and edges (relationships) in the graph database to support subsequent fast queries.

[0047] The content retrieval and interface module connects to the graph indexing construction module. This module provides query services to users. It receives query requests entered by users in a specific format, parses and converts these requests into formalized graph query schema statements that can be executed in the graph database. Subsequently, this module drives the graph database to execute the query statement, matches data that meets the conditions from the index built by the graph indexing construction module, and returns the processed search results to the user or upper-layer applications.

[0048] See attached document Figure 2 , Figure 2 This is a flowchart of a semantic indexing and content retrieval method for remote sensing time-series data according to an embodiment of the present invention. The method may include the following steps:

[0049] S301: Acquire raw remote sensing time-series images covering the target area and target time period, and perform preprocessing such as radiometric calibration, atmospheric correction, and geometric correction on the image data to generate a standardized numerical time-series dataset with accurate pixel alignment.

[0050] S302: For each pixel in the dataset, extract its numerical time series, and transform the numerical time series into a compact phenological primitive sequence containing morphological information through key point detection, phenological segment division and four-dimensional feature encoding.

[0051] S303. For each pixel, obtain its corresponding spatial context information and select a matching grammar rule from the phenological grammar knowledge base based on this information. Use the selected rule to parse the phenological primitive sequence of the pixel to generate a structured temporal relation graph containing entities and relationships.

[0052] S304 aggregates the temporal relationship graphs generated from all pixels and injects them into a global graph database. Within the graph database, indexes are created for the nodes (entities) and edges (relationships) of the graph, forming a unified graph index library that supports efficient querying.

[0053] S305: Receive the user's query request, parse the request, and convert it into a formalized graph query pattern. Execute the graph query pattern in the graph index, match all subgraph structures that meet the conditions, and extract the target information as the final retrieval result.

[0054] The steps of the method of the present invention will be described in detail below.

[0055] In step S301, a set of original remote sensing image data covering a preset geographical range and time span is first acquired.

[0056] In this embodiment, the data used are surface reflectance products from Landsat or Sentinel-2 satellites, which provide standardized data that has undergone preliminary radiometric calibration and atmospheric correction.

[0057] After acquiring the data, a quality assessment is performed on each image in the dataset. Using the quality assessment bands inherent to the image product, a quality mask is generated. This mask identifies and marks invalid pixels caused by interference from clouds, cloud shadows, snow, or other atmospheric conditions. This quality mask is applied to the corresponding image to ensure that only high-quality pixel values ​​are included in subsequent calculations.

[0058] On the quality-controlled image, a vegetation index is calculated for each valid pixel. This embodiment uses the Normalized Difference Vegetation Index (NDVI) as a physical quantity characterizing the dynamics of surface vegetation. Its calculation formula is as follows:

[0059] ;

[0060] in, For the surface reflectivity in the near-infrared band, This represents the surface reflectance in the red band. Performing this calculation on all images in the time-series data generates a raw NDVI time series for each pixel. Due to the application of quality masks, this raw series is irregular in the time dimension and contains missing data points.

[0061] To construct a time series that is temporally uniform and morphologically smooth, the original NDVI time series needs to be reconstructed and smoothed using filtering. This embodiment employs the Savitzky-Golay (SG) filtering algorithm to accomplish this task. This algorithm slides a fixed-size window across the time series and applies a low-order polynomial to the data points within the window for local least-squares fitting. This process effectively filters out high-frequency noise and fills in missing data points caused by cloud cover or other factors using the fitted values. The final output of this step is for each pixel... Generate a numerical time series with equal time intervals and a smooth curve shape. This sequence will be used as input data for step S302.

[0062] In step S302, the data is received as a single pixel. Generated smooth numerical time series This continuous numerical sequence is then transformed into a discrete sequence of phenological primitives containing structured morphological information. .

[0063] This step first involves key point detection to identify the time series. The key point marks a significant turning point in the phenological process. In this embodiment, the key point is determined by calculating the first and second derivatives of the time series, as follows:

[0064] For the input smooth numerical time series First, its first derivative sequence is calculated using the numerical difference method. and second derivative sequence .because It is a discrete time-point sequence, and its derivative is approximated by the following method:

[0065] For time points Its first derivative Calculated using the central difference formula:

[0066] ;

[0067] For the endpoints of the sequence, forward or backward difference is used.

[0068] Its second derivative Calculate by differentiating the first derivative or by directly using the following formula:

[0069] ;

[0070] in This represents a fixed time step for the time series.

[0071] After obtaining the first and second derivative sequences, different types of keypoints are identified by setting the following rules:

[0072] Local maxima: at time point At a certain point, if the first derivative undergoes a sign change from positive to negative, then the condition is satisfied. Then the time point This point is marked as a local maximum. It typically corresponds to the peak value of phenological growth.

[0073] Local minimum point: at time point At a certain point, if the first derivative undergoes a sign change from negative to positive, then the condition is satisfied. Then the time point This point is marked as a local minimum. It typically corresponds to the end of one phenological cycle and the beginning of the next.

[0074] Maximum growth rate point: at a certain time point At a certain point, if the second derivative undergoes a sign change from positive to negative, then the condition is satisfied. and Then the time point This point is marked as the point of maximum growth rate. This point corresponds to the first derivative. The peak value represents the moment when vegetation grows fastest.

[0075] Maximum decay rate point: at time point At a certain point, if the second derivative undergoes a sign change from negative to positive, then the condition is satisfied. and Then the time point This point is marked as the point of maximum decay rate. This point corresponds to the first derivative. The trough value represents the moment when vegetation withers or decays most rapidly.

[0076] The timestamps of all key points identified by the above rules are aggregated and arranged in chronological order to form a key point set. .

[0077] Based on the identified key point set For numerical time series Divide the phenological phases into segments. Output ordered timestamps. As a partitioning index, directly use the time series It is divided into an ordered set consisting of phenological segments that are connected end to end. Each paragraph Corresponding to time interval The sequence data fragment within.

[0078] Next, for each phenological section Perform four-dimensional feature encoding and convert it into a standardized phenological base. The phenological primitive is a quadruple, encoded by a function. generate:

[0079] ;

[0080] The four components are calculated as follows:

[0081] Trend Components This component describes the phenological section. The overall direction of change. It is determined by comparing the end points of paragraphs. Determine, and its calculation formula is:

[0082] ;

[0083] The calculation results (+1, 0, -1) are respectively mapped to a predefined set of discrete symbols, such as {Growth, Stable, Decline}.

[0084] Curvature component : This component describes the macroscopic concave and convex shape of the phenological paragraph . It is determined by calculating the sign of the cumulative deviation between the actual curve of the paragraph and the straight line connecting its two ends. Its calculation formula is:

[0085] ;

[0086] The calculation results (+1, 0, -1) are respectively mapped to a predefined set of discrete symbols, such as {Upward convex, Linear, Downward concave}.

[0087] Relative duration component This component describes the relative length of the duration of the phenological paragraph . The calculation process is as follows:

[0088] First, calculate its absolute duration .

[0089] Then, through a quantization function convert this absolute duration into a discrete symbol.

[0090] The threshold of this quantization function is determined based on the distribution characteristics (e.g., quartiles) obtained from the statistical analysis of the absolute durations of all pixels and all phenological paragraphs in the study area. This process is expressed as [[ID=4​​​​​​​​​​​​​​​​​​​​​​​​​​

[0094] Relative amplitude components This component describes the phenological section. The relative magnitude of the range of values ​​within the interval. Its calculation process is similar to that of the relative duration component.

[0095] First, calculate its absolute amplitude. .

[0096] Through a quantization function The absolute magnitude is converted into a discrete symbol.

[0097] The threshold of this quantization function is also determined based on the distribution characteristics obtained from statistical analysis of the absolute amplitudes of all phenological segments within the study area. This process represents... Its output relative amplitude components It is one of the predefined symbol sets, such as {low, medium, high}.

[0098] After encoding all phenological segments, the generated phenological base will be... Arranged according to their natural chronological order, the phenological primitive sequence of that pixel is ultimately formed. .

[0099] See attached document Figure 3 , Figure 3 This is a schematic diagram of a spatial adaptive grammar parsing process according to an embodiment of the present invention. In step S303, the phenological primitive sequence generated for a single pixel from step S302 is received. Based on the geographic environment of the pixel, the symbol sequence is parsed into a structured temporal relationship map. .

[0100] First, obtain the spatial context information corresponding to the current pixel. This information is extracted by spatially overlaying the geographic coordinates (latitude and longitude) of pixels with one or more pre-established prior geographic information layers in vector or raster form.

[0101] In this embodiment, the geographic information layers used include agricultural climate zoning maps, soil type distribution maps, and land use / land cover status maps. By querying these layers, pixels can be... Get an attribute vector, for example ={Climate zone: warm temperate semi-humid zone, Soil type: brown soil, Land use: arable land}.

[0102] At the same time, the system maintains a global phenological grammar knowledge base. This knowledge base stores multiple sets of grammar rules. Each set of rules corresponds to a specific spatial context.

[0103] In this embodiment, the knowledge base is organized as a lookup table with combinations of spatial context attributes as keys and subsets of grammar rules as values.

[0104] Phenological Grammar Knowledge Base The construction of [the object] can be achieved in, but is not limited to, the following two ways:

[0105] First, there is the expert-knowledge-based construction method. This method invites experts in agriculture, forestry, ecology, and other fields to define typical phenological primitive sequence patterns for different spatial contexts (such as a certain crop in a specific climate zone) based on their professional knowledge. For example, an agricultural expert can define the typical pattern corresponding to "winter wheat in the North China Plain" as follows: a primitive representing the overwintering period ("long duration, low amplitude, linear, stable"), followed by a primitive representing the greening and growth period ("medium duration, medium amplitude, concave, increasing"), and then a primitive representing the ripening and decay period ("medium duration, medium amplitude, convex, decreasing"). These expert-defined patterns are compiled into grammatical production rules and stored in a knowledge base.

[0106] Secondly, there is the data-driven automatic learning method. This method utilizes a training sample set with precise labels to automatically learn grammar rules through grammar induction or other machine learning algorithms. Specifically, it collects a large number of samples of known land cover types and their corresponding time-series data. After converting these time-series data into phenological primitive sequences, the algorithm automatically discovers frequently occurring sequence patterns that are highly correlated with specific land cover types (in a specific spatial context) and automatically generates corresponding grammar rules. This method can uncover complex patterns that even experts may not be able to summarize and can quickly adapt to new data.

[0107] In this embodiment, a hybrid approach is adopted, which uses expert knowledge to construct basic rules and then uses data-driven methods to expand and optimize the rule base, in order to ensure the accuracy and completeness of the knowledge base.

[0108] Based on the acquired spatial context information The system executes an adaptive selection function. To dynamically select the subset of matching grammar rules. This function will convert the pixel As a query condition, in the phenological grammar knowledge base The matching is performed to retrieve the most suitable subset of rules, denoted as... .

[0109] In this embodiment, the adaptive selection function The specific execution process is implemented as a multi-stage search strategy based on preset priorities to ensure that the input spatial context is always found. Find the most suitable subset of grammar rules. The process is as follows:

[0110] Phase 1: Exact Match Search

[0111] The function first takes the spatial context information of the input pixels. (For example, a set of attributes containing {land use: arable land, soil type: brown soil}) serves as a complete, composite query key. It uses this composite key in the phenological grammar knowledge base. The search is performed to locate an entry that perfectly matches the composite key. If an entry is found, its associated subset of grammar rules is... The result is returned directly, and the search process ends. This stage aims to apply the most specific and specialized grammar rules.

[0112] Phase Two: Priority Rollback Matching

[0113] If no exact match is found in the first stage, the function proceeds to this stage. This stage relies on a predefined list of spatial context attribute priorities. For example, this list could be defined as [land use, agroclimatic zoning, soil type], indicating that "land use" is the highest priority attribute and "soil type" is the lowest priority attribute.

[0114] The function will use this priority list, starting with the lowest priority attribute and gradually increasing the search criteria. Specifically:

[0115] 1. Remove the attribute with the lowest priority from the composite query key. In the example above, "soil type" is removed first, forming a new, more general query key {land use: arable land}.

[0116] 2. Using this new, simplified query key, repeat the exact match lookup from the first stage.

[0117] 3. If a match is found, return the corresponding subset of grammar rules, and the search process ends.

[0118] 4. If no match is found, and there are still removable attributes in the query key, continue removing the lowest priority attribute and repeat steps 2 and 3. This iterative process will continue until only the highest priority attribute remains.

[0119] Phase 3: Application of Default Rule Sets

[0120] If all priority backoff attempts in the second phase fail and no matching subset of grammar rules is found, the function executes this final phase. In this phase, the function retrieves data from the phenological grammar knowledge base. Retrieves and returns a subset of grammar rules that are pre-marked as "default". This default rule set contains a series of the most general and universal grammar rules. They do not depend on any specific spatial context and are designed to provide a basic and general semantic parsing scheme for pixels that cannot be specifically matched.

[0121] By executing the above three stages sequentially, the adaptive selection function ensures that for each pixel, based on its available spatial context information, one and only one most suitable subset of grammar rules is determined for subsequent parsing steps in an order from most specialized to most generalized.

[0122] Selected subset of grammar rules It contains a series of production rules, each rule A phenological primitive sequence pattern was defined for each. To generate actions from a graph The mapping. For example, a rule for identifying a typical single-crop-a-year crop can be defined as:

[0123] ;

[0124] Among them, the pattern to be matched It is a specific phenological element sequence, such as (<growth, concave, medium, high>, <decline, convex, medium, high>), which morphologically represents a complete growth-decline cycle.

[0125] Corresponding graph generation action The specific graph fragment to be generated is defined. In this example, the node set... Can be defined as , Phenological stages, name: growing season); boundary set Can be defined as v2,type: experience)}.

[0126] The grammar parsing engine selects a subset of grammar rules. Phenological primitive sequence applied to this pixel The engine uses a sequence matching algorithm to... Search and Patterns of each internal rule The matching subsequence. Once a match is found, the system executes the rule corresponding to that subsequence. Action: Instantiate a temporal relationship graph for this pixel. Excerpt.

[0127] The spatial adaptability of this method is reflected in the fact that, for completely identical phenological primitive sequences... If the spatial contexts of the pixels they belong to are different (e.g., one is in a cultivated area and the other is in a natural grass area), the system will select different subsets of grammar rules for parsing. For the natural grass area, the corresponding rules may use the same pattern. Parsing into different entities, such as Land cover (name: natural pasture). In this way, by introducing spatial context as the basis for selecting grammar rules, cases with similar temporal curves but completely different land cover types are effectively distinguished.

[0128] Generated temporal relationship map Relations (edges) in a graph can also carry attributes such as weights or labels to describe the strength or nature of the association. These attributes can be designed as descriptors that carry attributes, used to quantitatively or qualitatively describe the characteristics of the relationship itself, transforming the graph from a purely topological structure into a knowledge base containing depth information. The generation and assignment of these attributes occur during the graph generation process. It is done according to specific rules and calculation results.

[0129] Specifically, when a grammar rule is successfully matched, its corresponding Actions specify not only the source node, target node, and relationship type to be created, but also one or more attributes (key-value pairs) to be assigned to the relationship. These attribute values ​​primarily come from the following two aspects:

[0130] 1. Quantitative characteristics directly mapped from phenological primitives: The quantitative indicators of phenological segments calculated in step S302 can be directly used as attribute values ​​of relationships. For example, when identifying a land cover state transitioning from "normal growth" to "stressed," the absolute amplitude of the phenological segment that triggered the identification can be used. (e.g., the magnitude of the NDVI decrease) and absolute duration (Duration of coercion) is an attribute of this state transition relationship. A specific instance of the relationship could be: (source: node A, target: node B, type: transition, properties: {ndvi_change: -0.4, duration_days: 60}).

[0131] 2. Derived attributes assigned based on context or rule logic: Attribute values ​​can also be derived from the built-in logic of grammatical rules or external information. For example, when identifying a "land use change" (such as from "forest land" to "construction land"), this change relationship can be assigned the following attributes:

[0132] transition_date: The value of this attribute is set to the timestamp of the key point corresponding to this change. .

[0133] severity: The value of this attribute is based on the magnitude of change. A qualitative description obtained through a threshold judgment function, such as {severity: 'severe'}.

[0134] confidence: This attribute can be a preset value in the grammar rule (such as 0.9) to indicate the confidence level of this rule in recognizing this pattern in the current spatial context.

[0135] By attaching these specific, quantifiable attributes to relationships, this invention enables subsequent queries to not only retrieve "whether a certain relationship exists," but also to filter and analyze the "strength," "time," and "nature" of the relationship. For example, a user can initiate a query: "Retrieve all forest-to-grassland conversion events that occurred after 2024, with an NDVI decrease of more than 0.3 and a confidence level higher than 0.85."

[0136] In step S304, discrete temporal relationship maps generated for all pixels within the study area are received from step S303. The collection integrates and merges these independent, pixel-level map fragments into a unified, indexed, and optimized global temporal relationship map database covering the entire study area.

[0137] This embodiment uses an attribute graph as the underlying data model. In this model, the graph database contains nodes, relationships, labels, and attributes.

[0138] Nodes: Represent entities such as land cover types, phenological stages, or specific events. Each node is assigned one or more labels (e.g., land cover, event) to identify its category and may contain multiple key-value pairs of attributes (e.g., name: 'one crop per year', pixel_id: 12345).

[0139] Relationship: Represents a directed connection between nodes. Each relationship has one and only one type (e.g., experience, transformation) and may contain multiple attributes (e.g., transition_date:'2024-08-15', confidence:0.9).

[0140] The construction of the atlas index library is performed through a data injection and fusion procedure. This procedure processes the temporal relationship atlas generated for each pixel one by one. Fragment. For each node in a fragment, the program executes a "lookup or create and update" logic.

[0141] The specific execution process of this logic is as follows: The program first determines whether a node that is semantically equivalent to the currently processed node already exists in the global graph database. The equivalence criterion is based on a preset rule set, which comprehensively considers the node's label, key attributes, and spatial proximity. For example, two nodes from two adjacent pixels are considered semantically equivalent if both their label is "land feature" and both their name attribute is "one-crop-per-year crop," meaning they jointly describe the same real-world plot.

[0142] If the determination result indicates that an equivalent node already exists in the global graph, a "fusion" operation is performed. This operation updates the information of the current pixel onto the existing global node. Specifically, the attributes of the global node are aggregated and updated; for example, a pixel_count attribute is incremented by one, and its geographic extent attribute is expanded to include the coordinates of the current pixel. Simultaneously, relationships within the current graph fragment are reconnected to this existing global node.

[0143] If the determination result indicates that no equivalent node exists in the global graph, the "create" operation is executed. The program will create a new node in the global graph database and completely copy the node attributes from the current graph fragment to this new node. Simultaneously, the relationships in the graph fragment will also be created and connected to this new node.

[0144] By repeatedly executing the above injection and fusion logic on all pixel-centric map fragments, the discrete, pixel-centric map is reconstructed into a coherent global map centered on real geographic entities.

[0145] After data injection and fusion are completed, the system performs index building operations on the global graph database to improve subsequent query performance. Indexes are created on attributes in nodes and relationships that are frequently used as query filters. In this embodiment, the system creates indexes for at least the following attributes:

[0146] Create an index for the labels of all nodes to quickly filter all nodes in a specific category.

[0147] Create an index for the name attribute of the feature label node to quickly locate nodes of a specific feature type.

[0148] Create an index for the type of all relationships (edges) to quickly find associations of a specific type.

[0149] Create indexes for relationships with time attributes (such as transition_date) to support efficient time range queries.

[0150] Create a spatial index for the geographic coordinates attribute of the node to support efficient geospatial range queries.

[0151] The final output of this step is a single, complete, and highly indexed global temporal relationship graph database, which provides the data foundation and performance guarantee for the structured semantic retrieval in step S305.

[0152] See attached document Figure 4 , Figure 4 This is a schematic diagram of a structured semantic retrieval process according to an embodiment of the present invention. In step S305, a query request input by a user in a specific structured format is received, converted into a formal query statement that can be executed on the graph index library, and finally the exact matching retrieval result is returned.

[0153] This step first receives a structured query request following a predefined pattern. This request breaks down the user's query intent into the basic building blocks of the graph: nodes, relations, and attribute constraints imposed upon them. The content of a request specifies the graph pattern to be retrieved, for example, declaring a start node labeled "feature" with a "name" attribute of "woodland"; a target node labeled "feature" with a "name" attribute of "construction land"; and a directed relation of type "transformation" connecting the start and target nodes.

[0154] In addition to defining the topology of the graph schema, this structured request includes a series of constraints. These constraints are used to filter element attributes in the matching schema. For example, a constraint might require that the "transition_date" attribute value of the aforementioned "transform" relation must be later than a specific date, and that the "confidence" attribute value of the relation must be greater than a specific value. Finally, the request explicitly specifies which elements of the graph schema the query results should return.

[0155] Upon receiving the structured request, a query parsing and transformation module is responsible for converting it into a formal graph query statement that can be directly executed in the graph database. This transformation process follows a deterministic mapping rule. The graph schema defined in the request is converted into pattern matching clauses in the graph query language; the constraints defined in the request are converted into attribute filtering clauses in the query language; and the return elements specified in the request are converted into clauses defining the output results in the query language.

[0156] The generated graph query statement is submitted to the global temporal relation graph database constructed in step S304 for execution. Upon receiving the statement, the graph database's query engine efficiently executes the query plan using pre-built indexes. It quickly locates all candidate node sets that meet the conditions using indexes on node labels and attributes. Subsequently, it searches for paths that satisfy the join pattern among these candidate nodes using indexes on relation types. Finally, it filters the relation attributes using filter clauses to select all subgraph matching instances that fully satisfy the query conditions.

[0157] After the query is executed, the graph database returns a matching result set. Each record in the result set corresponds to a subgraph instance found in the global graph that matches the query pattern. This result set is then returned to a result processing module. This module is responsible for formatting the raw graph data into a form that is easy for users to understand and use. For example, it can organize the returned nodes and their attributes into a table, or extract the geographic coordinates associated with all matching "feature" nodes and aggregate them into a vector geographic data object for visualization on a map. The final processed result is returned to the user or upper-level application, thus completing the entire semantic retrieval process.

[0158] 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 semantic indexing and content retrieval method for remote sensing time-series data, characterized in that, Includes the following steps: Obtain the time series of pixel values ​​from remote sensing time series data; The numerical time series is converted into a phenological primitive sequence; Based on the spatial context corresponding to the pixel, a matching grammar rule is selected from a preset phenological grammar knowledge base; Based on the grammar rules, the phenological primitive sequence is parsed to generate a temporal relationship map representing the semantic information of the pixels; The generated temporal relationship graph is constructed into a graph index library; A graph query pattern is generated based on the user's input query request, and then matched in the graph index library to obtain search results; The step of converting the numerical time series into a phenological primitive series includes: Key point detection is performed on the numerical time series to identify phenological turning points; The numerical time series is divided into at least one phenological segment using the phenological inflection point. For each phenological segment, encoding is performed from four dimensions: trend, curvature, relative duration, and relative amplitude, to generate phenological primitives corresponding to the phenological segment, thus forming the phenological primitive sequence; The grammar rules define the mapping relationship between phenological primitive sequence patterns and actions for generating temporal relationship graphs; The temporal relationship graph includes entities used to describe semantic concepts, and relationships used to describe the associations between the entities.

2. The semantic indexing and content retrieval method for remote sensing time-series data according to claim 1, characterized in that, The spatial context is attribute information extracted from at least one prior geographic information layer based on the geographic coordinates of the pixel.

3. The semantic indexing and content retrieval method for remote sensing time-series data according to claim 1, characterized in that, The entity includes at least one of the following: a land feature, a phenological stage, or an event; the relationship is used to describe the association between the land feature, the phenological stage, and the event.

4. The semantic indexing and content retrieval method for remote sensing time-series data according to claim 1, characterized in that, The graph index library is built on a graph database and includes: Entity indexes used to index entities in the time-series relation graph; Adjacency index used to index the relationships in the time-series relationship graph.

5. The semantic indexing and content retrieval method for remote sensing time-series data according to claim 1, characterized in that, The step of generating a graph query pattern based on the user's input query request includes: The semantic concepts and logical constraints contained in the query request are converted into query limits for entities, relationships and their topological structures in the temporal relationship graph.

6. The semantic indexing and content retrieval method for remote sensing time-series data according to claim 1, characterized in that, The same phenological primitive sequence, when associated with different spatial contexts, is parsed into temporal relation graphs containing different semantic information through the adaptive selection of the grammar rules.

7. The semantic indexing and content retrieval method for remote sensing time-series data according to claim 1, characterized in that, The step of generating the time-series relation graph further includes: Add weights or labels to the relationships to represent the associated attributes, so as to support complex queries with attribute constraints in the graph index.

8. A semantic indexing and content retrieval system for remote sensing time-series data, used to perform the method as described in any one of claims 1-7, characterized in that, include: The symbolization conversion module is used to obtain the numerical time series of pixels in remote sensing time series data and convert them into phenological primitive sequences; An adaptive grammar parsing module is used to select matching grammar rules from a preset phenological grammar knowledge base according to the spatial context corresponding to the pixel, and parse the phenological primitive sequence based on the grammar rules to generate a temporal relation graph representing the semantic information of the pixel. The graph index construction module is used to construct the generated time-series relationship graph into a graph index library; The content retrieval module is used to generate a graph query pattern based on the user's input query request and match it in the graph index library to obtain the retrieval results.

Images