A remote sensing algorithm online operation intelligent recommendation method and system for public interest litigation

Abstract

The application belongs to the field of remote sensing algorithm process modeling and recommendation, and discloses a remote sensing algorithm online operation intelligent recommendation method for public prosecution of public interest litigation, acquires a certain number of remote sensing algorithm processes of environmental resource public interest litigation, and performs data cleaning and semantic classification on nodes; the standardized remote sensing algorithm process is matrixed to obtain a feature matrix; the feature matrix is subjected to an independent path extraction algorithm to obtain a plurality of independent path sequences; a prefix of the independent path sequence is subjected to frequent item mining to obtain a prefix part for training, and a tail node is used as a label; long-term features are extracted through a pyramid RNN model, and a random gradient descent algorithm is used to train a neural network to obtain a training model. The application simplifies the recommendation of remote sensing algorithms for environmental resource public interest litigation, converts a graph recommendation task into a sequence recommendation task, simplifies the knowledge demand of algorithm modeling personnel on a complex graph structure, and improves the recommendation efficiency of a recommendation system.

Description

A Remote Sensing Algorithm-Based Online Intelligent Recommendation Method and System for Public Interest Litigation by Procuratorates Technical Field

[0001] This invention belongs to, but is not limited to, the field of remote sensing algorithm process modeling and recommendation, and particularly relates to an intelligent recommendation method and system for online operation of remote sensing algorithms for public interest litigation by procuratorates. Background Technology

[0002] Digital reform of the procuratorate is an important measure to strengthen legal supervision in the new era. As a key link supporting procuratorial case handling, forensic identification and other procuratorial technical work provide indispensable evidence and technical support for case handling. Currently, procuratorial organs nationwide have basically established an "integrated air-ground" technical support system for public interest litigation, centered on technologies such as satellite remote sensing, drones, rapid detection, and big data analysis, which has significantly improved the efficiency of handling public interest litigation cases.

[0003] However, in the practical application of remote sensing technology in handling public interest litigation cases, two major challenges are encountered: First, the operational bottleneck for users. Prosecutors, as non-professional users, find it difficult to independently and efficiently utilize existing remote sensing application systems designed for technical professionals to intelligently mine and analyze multi-source, multi-dimensional, and heterogeneous data. Second, the dilemma of algorithm selection. The diversity of remote sensing algorithms in multi-task and single-task scenarios during case handling means that the merits of various algorithms are not absolute and are influenced by many factors, such as the characteristics of the remote sensing data source, including the spatial and spectral resolution of the imagery. That is, the algorithm's performance will differ depending on the input data source. This necessitates constructing corresponding sample datasets for different data sources during algorithm evaluation, while also considering situations where external conditions prevent the application of suitable remote sensing images for the optimal algorithm in the research area. Therefore, it is necessary to research a multi-source algorithm intelligent recommendation mechanism for the same task, providing prosecutors with a convenient and effective algorithm selection solution.

[0004] While recommender systems are widely used in academia and industry, their application to business process recommendations, particularly in the integration of digital prosecution and remote sensing technologies, remains a relatively new area. Traditional graph similarity-based methods, such as those based on graph edit distance, can achieve accurate matching, but their computational complexity increases exponentially with the number of nodes, failing to meet the real-time recommendation needs of public interest litigation scenarios. Other process similarity calculation methods are optimized based on structural similarity, but still lack some important information, and their accuracy and efficiency need further improvement. Furthermore, when faced with large amounts of prosecutorial text information, recommender algorithms need to capture complete semantic information. Existing sequence-based recommender methods generally employ traditional recurrent neural network architectures such as LSTM, which, while capturing temporal dependencies between nodes rather than continuous semantic segments, have significant limitations in modeling capabilities. These methods often treat process nodes as discrete, independent events, failing to effectively model semantic segment relationships across time steps. Moreover, these methods lack context awareness of process prefix sequences.

[0005] The closest existing technology is the patent "Remote Sensing Process Recommendation Method Based on Logical Structure and Attribute Features" (CN116226220A) and the corresponding algorithm framework proposed at the SPIE conference. Both methods parse the existing process's structure tree and attribute set, calculate path similarity, and then recommend the next node or complete process in a Top-N manner. This approach, like the one in this invention, belongs to the "workflow-level" algorithm recommendation category, but it mainly relies on static logical structure matching and lacks in-depth extraction of node semantics, running status, and real-time business scenarios. Furthermore, it does not utilize long sequence modeling capabilities, resulting in limited prediction effectiveness for complex branches, loops, or dynamically changing process prefixes.

[0006] The existing solutions have two major technical problems: First, at the algorithm level, there is a lack of end-to-end learning of neural networks to address the temporal dependence of processes, which leads to a decrease in recommendation accuracy when nodes have similar semantics but large topological differences, making it difficult to support the high-confidence, low-latency online reasoning requirements of remote sensing monitoring for public interest litigation; Second, at the platform level, continuous incremental data cleaning, semantic classification, and online feedback loops are not integrated, which cannot solve the pain points such as cold start, concept drift, and sparse long-tail nodes that occur after the rapid expansion of the remote sensing service database, making it difficult to form an interpretable and adaptive intelligent recommendation chain. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides an intelligent recommendation method for online operation of remote sensing algorithms for public interest litigation by procuratorates.

[0008] This invention is implemented as follows: an online intelligent recommendation method for remote sensing algorithms used in public interest litigation by procuratorates, the algorithm comprising the following steps:

[0009] a. Data on remote sensing algorithm flowcharts for environmental resource public interest litigation were collected, primarily from domestic and international remote sensing journals. A focus was placed on papers from the past 20 years covering sub-fields such as water pollution, land resources, and water area changes, from which flowcharts were extracted. Data cleaning and semantic classification were performed to generate standardized flowcharts. Due to the lack of standardization in the original flowchart representations, there were slight differences in the descriptions of various flowcharts. This paper standardizes the flowchart structure. First, loop structures in the dataset were removed. Loop structures in remote sensing algorithm flowcharts often represent repetitive processes, do not affect the overall logic, and account for a very small proportion. Second, explicit representations of parallel and selection structures were added, namely, two new flowchart nodes: "and" and "or." Only these two types of flowchart nodes have an in-degree or out-degree greater than 1. After organizing the logical structure, the semantics of the flowchart nodes were clustered and deduplicated to avoid the same flowchart node being divided into multiple categories due to differences in scholars' expressions. Finally, the flowchart was abstracted into a directed graph.

[0010] b. The standardized process is matrixed to obtain a feature matrix; the directed graph is transformed into an adjacency matrix. After abstracting the flowchart into a directed graph, it is further transformed into an adjacency matrix for easier subsequent program processing. Each directed graph can be represented by a two-dimensional matrix. The diagonal of the matrix stores the node numbers (i.e., the unique identifiers of the process nodes), and the other positions of the matrix represent the information of the directed edges. 1 indicates that the process node in the current row has an edge pointing to the process node in the current column, and 0 indicates that there is no process pointing relationship between the two process nodes.

[0011] c. Perform independent path extraction on the feature matrix to obtain multiple independent path sequences; AND nodes with an in-degree greater than 1 are AND-joined, and AND nodes with an out-degree greater than 1 are AND-split; similarly, OR nodes with an in-degree greater than 1 are OR-joined, and OR nodes with an out-degree greater than 1 are OR-split. The path extraction logic is as follows:

[0012] (1) Ordinary node: Add its successor node to the end of the current path. (2) Control node: If the current path already contains all its direct predecessor nodes, add the successor node to the path; otherwise, backtrack to the unprocessed predecessor node and update the path; the and-split node performs a full permutation of all its successor nodes, generates the corresponding new path and enqueues it; the or-join node adds its unique successor node to the current path; the or-split node traverses all its successor nodes, generates new paths in turn and enqueues them; (3) End node: Collect the current path, remove the intermediate control node information and store it in an independent path table.

[0013] d. Perform frequent term mining on the independent path sequences described in the previous step, using the prefix as training input and the tail node as the label; the lower layers of the network process fine-grained sequence information, while the higher layers process coarse-grained sequence information. The output of the lower-level RNN serves as the feature of that layer and is also upsampled to match the time step length of the input to the higher-level RNN. The upsampled lower-level features are concatenated or added to the input of the higher-level RNN before being fed back into the higher-level RNN.

[0014] e. Use a pyramid recurrent neural network model to train a large number of training samples to obtain a fully trained recommendation model;

[0015] f. Based on the recommendation model, generate subsequent process node recommendation results for the input remote sensing process prefix. Further, the specific process of creating and standardizing the remote sensing monitoring algorithm dataset in step (1) is as follows: First, perform semantic classification on the process nodes in the algorithm dataset and assign a unique node identifier to each type of node; replace the original process data nodes with the new node identifiers to obtain the standard dataset.

[0016] Furthermore, the specific steps for process matrixization in step (2) are as follows: the marked directed graph is converted into the form of an adjacency matrix to obtain the feature matrix. The position of 1 in the feature matrix indicates that there is a directed edge between the two corresponding nodes. The diagonal position in the feature matrix contains the unique identifier of the node.

[0017] Further, the specific process of the independent path extraction algorithm in step (3) for extracting independent paths from the feature matrix is ​​as follows: AND nodes with an in-degree greater than 1 are denoted as AND-split, and AND nodes with an out-degree greater than 1 are denoted as AND-join. Similarly, OR nodes with an in-degree greater than 1 are denoted as OR-join, and OR nodes with an out-degree greater than 1 are denoted as OR-split. If the current process node executes a normal process node, the subsequent process node will be inserted at the end; if the current process node executes a control node, the corresponding in-degree and out-degree will be calculated according to the node type. If the current node is AND-join, it is necessary to determine whether the current process sequence contains all directly in-degree nodes. If it contains all directly in-degree nodes, the subsequent node is directly inserted into the path and returned to the queue; otherwise, backtracking is performed to find the two previous nodes for judgment, and the new path is updated into the path, which determines the node to be judged and the insertion position in the next loop. If the current node is AND-split, it means that all process nodes with its out-degree need to be executed. Here, the subsequent nodes are directly permuted and combined using the full permutation algorithm and added to the path and enqueued respectively. If the current node is an OR-join node, it will only have one outgoing node, which can be directly added to the path and enqueued. If the current node is an OR-split node, it means that any of the subsequent nodes will be selected for execution. In this case, all subsequent nodes are traversed and added to the path and enqueued in turn. If the current node reaches the end node, the results need to be collected, and the intermediate flow control node information is removed and saved to a separate path table.

[0018] Furthermore, the specific steps of frequent item mining in step (4) are as follows: frequent item mining is performed on independent path sequences using a recursive algorithm, and the sequences are uniformly divided into process prefixes and subsequent nodes.

[0019] Furthermore, the pyramid RNN model in step (5) includes multi-scale feature extraction and feature fusion through operations such as upsampling and concatenation.

[0020] Another objective of this invention is to provide an intelligent recommendation system for online operation of remote sensing algorithms for public interest litigation, which implements the aforementioned intelligent recommendation method for online operation of remote sensing algorithms for public interest litigation. The system includes:

[0021] The user interface module serves two purposes. First, it provides system administrators with a convenient entry point, allowing them to import new processes and thus enable incremental process expansion. Second, it provides modelers with a highly interactive environment, facilitating their process modeling work.

[0022] The offline processing module is responsible for offline processing tasks of the original business process data.

[0023] The process library module is responsible for the data storage function of the entire recommendation system. It stores various types of data for the entire recommendation system, including raw process flow data, as well as sequence data and graph pair data obtained after offline processing modules.

[0024] The online recommendation module is responsible for completing the process recommendation task through recommendation algorithms.

[0025] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, the computer program being executed by the processor causing the processor to perform the steps of the intelligent recommendation method for online operation of remote sensing algorithms for public interest litigation by procuratorates.

[0026] Another objective of this invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the intelligent recommendation method for online operation of remote sensing algorithms for public interest litigation by procuratorates.

[0027] Another objective of this invention is to provide an information data processing terminal, which is used to realize the online operation of the intelligent recommendation system for remote sensing algorithms oriented towards public interest litigation by procuratorates.

[0028] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0029] First, this invention simplifies the recommendation algorithm for remote sensing in environmental and resource public interest litigation by transforming the graph recommendation task into a sequence recommendation task. This simplifies the knowledge requirements of algorithm modelers regarding complex graph structures and improves the recommendation efficiency of the recommendation system.

[0030] In this invention, the model training and testing phases are separated. Although the data preprocessing and offline training phases are relatively time-consuming, the testing and recommendation phase takes less time and can meet the needs of real-time recommendation.

[0031] This invention does not require a specific structure for the flowchart. The original algorithm flow can be re-expressed using a directed graph and then processed into data structures such as two-dimensional matrices for the algorithm to call, thus effectively avoiding the limitations of traditional methods.

[0032] Secondly, the expected benefits and commercial value of the technical solution of this invention after its transformation are mainly manifested in improving the efficiency of judicial case handling and reducing labor costs. Primarily targeting procuratorial organs nationwide, it can support the handling of thousands of environmental litigation cases.

[0033] This invention fills a technological gap in the industry both domestically and internationally: it is the first to combine public interest litigation by prosecutors, remote sensing algorithms, and algorithm recommendation technology. It addresses the specific problems inherent in judicial scenarios. A pyramid RNN model is used to process the workflow topology, and an incremental workflow library update mechanism is established.

[0034] The technical solution of this invention solves a long-standing technical problem that has remained unsolved: it overcomes the operational bottleneck for non-remote sensing professionals such as prosecutors. Through node semantic classification encoding and independent path sequence optimization, it achieves "natural language-like" interaction. Furthermore, by combining matrix feature fusion and an online feedback closed-loop mechanism, it transforms subjective experience-driven remote sensing algorithm selection into intelligent recommendation.

[0035] The technical solution of this invention overcomes technical bias: traditional methods consider graph editing distance to be irreplaceable, but this invention improves the recommendation efficiency of the algorithm while ensuring accuracy by using path serialization and pyramid RNN. Attached Figure Description

[0036] Figure 1 is a schematic diagram of a typical architecture of a process recommendation system;

[0037] Figure 2 is an example diagram of the original algorithm flow;

[0038] Figure 3 is an example diagram of the algorithm flow after adding a gateway node;

[0039] Figure 4 is an example diagram of a standardized algorithm flow;

[0040] Figure 5 is a schematic diagram of converting the algorithm flowchart in Figure 4 into an adjacency matrix;

[0041] Figure 6 shows the architecture of the next process node prediction model based on the fusion of long and short-term sequence features;

[0042] Figure 7 shows the Pyramid RNN architecture;

[0043] Figure 8 shows the structure of a self-attention network.

[0044] Figure 9 is a comparison of the time taken for a single recommendation using different methods;

[0045] Figure 10 shows a comparison of HR@K results for different Hit-Rate (HR). Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0047] This invention provides an online intelligent recommendation method for remote sensing algorithms used in public interest litigation by procuratorates. The algorithm includes the following steps:

[0048] (1) Obtain a certain number of remote sensing algorithm processes for public interest litigation concerning environmental resources. After data cleaning and semantic classification of nodes, a standard remote sensing algorithm process is obtained.

[0049] (2) The standardized remote sensing algorithm process is matrixed to obtain the feature matrix.

[0050] (3) The feature matrix is ​​processed by an independent path extraction algorithm to obtain multiple independent path sequences.

[0051] (4) Extract frequent terms from the prefix of the independent path sequence. The mined prefix part is used for training, and the last node is used as the label.

[0052] (5) A large number of training samples are used as input to the neural network. Long-term features are extracted through the pyramid RNN model. The neural network is trained using the stochastic gradient descent algorithm to obtain the training model. Subsequent process nodes are recommended by inputting remote sensing process prefixes.

[0053] The specific process of creating and standardizing the remote sensing monitoring algorithm dataset in step (1) is as follows: First, the process nodes in the algorithm dataset are semantically classified, and a unique node identifier is assigned to each type of node; the original process data nodes are replaced with new node identifiers to obtain a standard dataset.

[0054] The specific steps for process matrixing in step (2) are as follows: the marked directed graph is converted into the form of an adjacency matrix, which yields the feature matrix. The position of 1 in the feature matrix indicates that there is a directed edge between the two corresponding nodes. The diagonal position in the feature matrix contains the unique identifier of the node.

[0055] The specific process of the independent path extraction algorithm in step (3) for extracting independent paths from the feature matrix is ​​as follows: AND nodes with an in-degree greater than 1 are denoted as AND-split, and AND nodes with an out-degree greater than 1 are denoted as AND-join. Similarly, OR nodes with an in-degree greater than 1 are denoted as OR-join, and OR nodes with an out-degree greater than 1 are denoted as OR-split. If the current process node executes a normal process node, the subsequent process node will be inserted at the end; if the current process node executes a control node, the corresponding in-degree and out-degree will be calculated according to the node type. If the current node is AND-join, it is necessary to determine whether the current process sequence contains all directly in-degree nodes. If it contains all directly in-degree nodes, the subsequent node is directly inserted into the path and returned to the queue; otherwise, backtracking is performed to find the two previous nodes for judgment, and the new path is updated into the path, which determines the node to be judged and the insertion position in the next loop. If the current node is AND-split, it means that all process nodes with its out-degree need to be executed. Here, the subsequent nodes are directly permuted and combined using the full permutation algorithm and added to the path and enqueued respectively. If the current node is an OR-join node, it will only have one outgoing node, which can be directly added to the path and enqueued. If the current node is an OR-split node, it means that any of the subsequent nodes will be selected for execution. In this case, all subsequent nodes are traversed and added to the path and enqueued in turn. If the current node reaches the end node, the results need to be collected, and the intermediate flow control node information is removed and saved to a separate path table.

[0056] The specific steps of prefix extraction and frequent item mining in step (4) are as follows: frequent item mining is performed on independent path sequences through a recursive algorithm, and the sequences are uniformly divided into process prefixes and subsequent nodes.

[0057] The pyramid RNN model in step (5) includes multi-scale feature extraction and feature fusion through operations such as upsampling and concatenation.

[0058] Figure 1 illustrates the typical architecture of the remote sensing algorithm recommendation system for environmental and resource public interest litigation, mainly comprising four parts: a user interface module, an offline processing module, a process library module, and an online recommendation module. The user interface module plays a crucial interactive role in the system. On one hand, it provides system administrators with a convenient entry point, supporting the import of new processes to achieve incremental process expansion. On the other hand, this module provides modelers with a highly interactive environment, facilitating their process modeling work. The offline processing module is a key module in the system, primarily responsible for the offline processing of raw business process data. The process library module is responsible for the data storage function of the entire recommendation system. It stores various types of data for the entire recommendation system, including both raw process flow data and sequence data and graph pair data obtained after offline processing. The online recommendation module, as the core of the entire recommendation system, is responsible for completing the process recommendation task through the recommendation algorithm.

[0059] The method of this invention is used in an online recommendation module. Its core processing framework includes data standardization, independent path extraction, long-term feature extraction, short-term feature extraction, long- and short-term feature fusion, and classification. Specifically, it includes the following steps:

[0060] (1) Data standardization.

[0061] Figures 2 and 3 illustrate part of the data standardization process, starting with flowchart data cleaning and node semantic classification. Due to the lack of standardization in the original flowchart representations, there are slight differences in the descriptions of various flowcharts. This paper standardizes the flowchart structure. First, loop structures in the dataset are removed. Loop structures in remote sensing algorithm flowcharts often represent repetitive processes, do not affect the overall logic, and account for a very small proportion. Second, explicit representations of parallel and selection structures are added, i.e., two new types of flowchart nodes are added: `and` and `or`. Only these two types of flowchart nodes will have an in-degree or out-degree greater than 1.

[0062] The flowchart is abstracted as a directed graph. Building on the previous section, the flowchart is further abstracted into a directed graph representation. Here, the node number indicates the type of the flowchart node in the flowchart library, and the directed edges represent the flow direction of the flowchart nodes.

[0063] Figure 5 illustrates the transformation of a directed graph into an adjacency matrix. The diagonal of the matrix stores the node numbers (i.e., the unique identifiers of process nodes), while the other positions in the matrix represent information about directed edges. A 1 indicates that the process node in the current row has an edge pointing to the process node in the current column, and a 0 indicates that there is no process pointing relationship between the two process nodes.

[0064] (2) Frequent item mining based on independent paths.

[0065] Specifically, the algorithm employs a breadth-first search approach, transforming the directed graph adjacency matrix into independent execution sequences based on the semantics of the process nodes. Then, frequent item mining is performed on the extracted independent path sequences, ensuring continuity. For a process sequence [a,b,d,e], with a length of 4, frequent sequences starting with 'a' are first mined, resulting in [a][a,b][a,b,d][a,b,d,e]. Next, frequent sequences starting with 'b' are mined, resulting in [b][b,d][b,d,e], and so on, mining frequent sequences starting with 'd' and 'e'. Finally, meaningless process sequences of length 1 are removed, leaving all process sequences longer than 1 as the training dataset for the model.

[0066] (3) Long-term feature extraction.

[0067] A pyramid recurrent neural network is used to capture the multi-scale long-term features of the process sequence, and its architecture is shown in Figure 7. The multi-scale convolutional pyramid module can be divided into two parts: bottom-up one-dimensional convolutional layers and top-down upsampling layers. The convolution period of the one-dimensional convolutional layers is preset, and the kernel size of each layer is set to a multiple of the window length of the current layer to the window length of the subsequent layer, with the stride being equal to the kernel size. Finally, all one-dimensional convolutions have the same number of output channels to facilitate upsampling. The vector representation of the remote sensing process sequence after the i-th one-dimensional convolution is as follows:

[0068]

[0069] In the formula Indicates the remote sensing process sequence after the first... Vector representation after one-dimensional convolution of a layer.

[0070] The upsampling process starts with sampling from the highest level of the pyramid and performs feature fusion in the lateral connections. The formula below illustrates the generation process of the features at the i-th layer:

[0071]

[0072]

[0073] The feature vectors after passing through multi-scale convolutional modules are used as input to gated recurrent units (GRUs), and finally the outputs of the multi-layer GRUs are processed through weights. The weighted and concatenated vector representation used as a process prefix, where Depend on Obtained through the softmax function These are learnable parameters, and are used... Perform initialization. The formula is as follows:

[0074]

[0075]

[0076]

[0077] (4) Short-term feature extraction.

[0078] In remote sensing algorithm workflows, initialization processes typically follow remote sensing data. Many algorithm processes exhibit certain patterns in local proximity. Considering the direct impact of preceding processes on subsequent process nodes, the nearest l process nodes in the process sequence are extracted as subsequences. The positional encodings in the subsequences are fused, and a stacked multi-layer self-attention network is used to extract short-term sequence features. The structure of the self-attention network is shown in Figure 8.

[0079] Since ordinary self-attention is not sensitive to positional information in the sequence, positional encoding (PE) is added to incorporate positional information into the sequence vector. The positional encoding formula and the formula for the nearest-flow sequence vector with embedded positional information are shown below:

[0080]

[0081]

[0082] Mode middle, , represents the position index in the sequence. Index representing a dimension This represents the embedding dimension of the process node. In the formula, This represents the process sequence vector after short-term feature extraction.

[0083] Next, the input is fed into a multi-layer dot-product attention layer to obtain the relationship between different process nodes in the process sequence and to calculate the importance of each process node to the next process recommendation.

[0084] (5) Feature fusion.

[0085] Using the same method as the short-term feature extraction layer, an attention network is used to fuse the features of both. The long-term features are used as the query vector, and the short-term features are used as the key vector and value vector. The calculated result contains complete information about the long and short-term sequences.

[0086] Implementation Case: A randomly selected independent path [39,122,61,26,55,110,113,92] provides a water quality parameter inversion monitoring workflow. The specific workflow semantics are: [water quality parameter remote sensing image data, radiometric calibration, atmospheric correction, reflectance calculation, mathematical transformation, Pearson correlation analysis, screening of band features or combinations, and construction of a water quality parameter inversion model]. The actual direct subsequent node is 125, representing "water quality parameter inversion". The complete workflow corresponding to this independent path is as follows:

[0087] This method first acquires hyperspectral or multispectral image data of the water body to be measured through a multi-source remote sensing platform, and immediately performs preprocessing such as radiometric calibration and atmospheric correction after acquisition to remove sensor response differences and atmospheric scattering and absorption effects, thereby obtaining an accurate surface reflectance spectrum curve, laying the foundation for subsequent quantitative analysis.

[0088] In the feature extraction stage, Pearson correlation analysis is used to screen all bands and their combinations based on the spectral response characteristics of different water quality parameters (such as chlorophyll a, suspended solids, or DOM). Feature bands or band ratios that are highly correlated with the target parameters are selected, and then an inversion model based on machine learning algorithms such as linear regression, support vector machine, or random forest is constructed to achieve quantitative prediction of water quality parameters.

[0089] Finally, the trained inversion model is applied to the preprocessed full-frame image to calculate the concentration of water quality parameters pixel by pixel, and a spatiotemporal distribution map of water quality is generated by combining time-series images. Through statistical analysis and visualization, the location and time-period change patterns of water pollution points are revealed, providing an intuitive basis for accurate water environment monitoring and intelligent decision-making.

[0090] The recommended results of this invention for the above process are as follows:

[0091]

[0092] This invention provides an intelligent recommendation system for online operation of remote sensing algorithms for public interest litigation, implementing the aforementioned intelligent recommendation method for online operation of remote sensing algorithms for public interest litigation. The system includes:

[0093] The user interface module serves two purposes. First, it provides system administrators with a convenient entry point, allowing them to import new processes and thus enable incremental process expansion. Second, it provides modelers with a highly interactive environment, facilitating their process modeling work.

[0094] The offline processing module is responsible for offline processing tasks of the original business process data.

[0095] The process library module is responsible for the data storage function of the entire recommendation system. It stores various types of data for the entire recommendation system, including raw process flow data, as well as sequence data and graph pair data obtained after offline processing modules.

[0096] The online recommendation module is responsible for completing the process recommendation task through recommendation algorithms.

[0097] After the system starts, the user interface module first receives flowchart files imported by modelers or administrators. It then uses a front-end parser to pass the flowchart nodes and connection structures to the back-end service, displaying node attributes and connectivity relationships in real time on the page. The interface also provides node label editing, version rollback, and incremental import functions. When a new flowchart submission is detected, the back-end offline processing module is triggered to standardize and preprocess the new data.

[0098] The offline processing module sequentially completes loop removal, parallel and selection structure explicitation, and semantic clustering deduplication according to a predetermined process. It then maps the normalized directed graph into an adjacency matrix and a feature matrix, followed by independent path extraction and frequent item mining to generate a high-support path prefix-tail node mapping dataset. Subsequently, this module calls the pyramid recurrent neural network training subsystem to complete iterative model training based on a multi-scale feature fusion strategy, and writes the trained model parameters into the process library module.

[0099] The process library module acts as a persistence center, storing raw flowchart data, preprocessed sequence data, and graph pair data in structured tables. It also maintains the latest trained model and historical model versions. When receiving query requests from online modules, this module can quickly locate the corresponding prefix mapping table and model parameters, achieving millisecond-level read / write performance, and providing a complete data link for subsequent auditing, retraining, or topology analysis.

[0100] The online recommendation module receives process prefixes pushed by the user interface or external systems during runtime. It ensures orderly data delivery under high concurrency through a message queue, and then outputs several candidate subsequent nodes based on the pyramid recurrent neural network model in the process library, returning them to the caller. The module also records recommendation logs and supports a model feedback interface, so as to periodically call the offline processing module to evaluate and retrain the recommendation effect, thereby forming a closed-loop operation mechanism for continuous optimization.

[0101] The hardware environment of this invention is based on the Ubuntu 18.04.6 LTS operating system, with a 6-card 2080Ti 11GB graphics card and 128GB of internal storage. This invention is implemented using PyTorch with the Adam optimizer, a learning rate of 0.0001, a batch size of 128, a dropout rate of 0.2, a process embedding dimension of 32, a self-attention layer of 2, a short-term window size of 6, and a long-term sequence extraction window of [2, 4, 8]. Cross-validation is used to eliminate the influence of randomness in the distribution of the randomly partitioned dataset. During the experiment, one set is selected sequentially as the test set, and the other four sets are used as the training set. The average of the five experimental results is taken as the final experimental result. Two metrics are selected to evaluate the accuracy and efficiency of the recommendation. The specific metrics are as follows:

[0102] (1) HR@K: Hit-Rate (HR) calculates the score of the target process node among the K process nodes with the highest recommendation probability. The calculation method of HR@K is as follows:

[0103] (2)

[0104] In the formula, It is an indicator function that means 1 if the actual process node is in the recommended list of K process nodes, and 0 otherwise.

[0105] (2) Average time per recommendation: The time taken by the evaluation model for each recommendation in the test set, in milliseconds.

[0106] This invention uses LCS-based, GRU, Bi-LSTM, shallow LSTM-based algorithm, multi-scale CNN model, and ProcessTransformer model for comparative experiments. The results are as follows:

[0107]

[0108] This invention achieves optimal results on four metrics: HR@1, HR@2, HR@3, and HR@5, with values ​​of 0.8207, 0.9000, 0.9331, and 0.9530, respectively. In particular, it improves upon the suboptimal model by 0.47% and 0.18% on the accurate recommendation metrics HR@1 and HR@2, respectively. Its advantages likely stem from two aspects: the short-term sequence feature extraction layer enhances the sensitivity to neighboring process nodes, thereby improving the accuracy of the precise recommendation; and the multi-scale feature fusion strategy combined with an attention mechanism effectively captures the spatiotemporal correlation of process nodes.

[0109] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0110] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A remote sensing algorithm-based intelligent recommendation method for public interest litigation by procuratorates, characterized in that, Includes the following steps: a) Collect remote sensing algorithm flow data for environmental resource public interest litigation, perform loop removal, parallel node explicit selection, and semantic clustering to generate a standardized directed graph; b) Transform the standardized directed graph into a feature matrix and an adjacency matrix; c) Extract independent paths from the feature matrix, identify parallel merging and branching nodes, and generate multiple independent path sequences; d) Perform frequent item mining on the independent path sequences, uniformly divide them into flow prefixes and subsequent nodes, statistically analyze the correspondence between high-support prefixes and tail nodes, use the prefixes as training input, and use the tail nodes as labels. e) The recommendation model is trained based on the pyramid recurrent neural network model, and the low-level features are upsampled and fused with the high-level features; f) The recommendation model generates subsequent process node recommendation results for remote sensing process prefixes; In step a, semantic clustering and deduplication of process nodes is achieved by assigning unique node identifiers and replacing the original nodes; In step b, matrixization is achieved by filling the node identifiers in the diagonal of the matrix and setting the matrix element corresponding to the process pointer to one; In step c, when a parallel merging node with an in-degree greater than one is encountered, it is determined that all direct predecessors and subsequent paths have been included and backtracked for updating; when a parallel system node with an out-degree greater than one is encountered, a full permutation is generated to generate the path; In step d, frequent item mining is achieved by recursively traversing the path sequence, splitting the path, and counting the occurrence frequency to obtain a high-support prefix set; In step e, the pyramid recurrent neural network model is set with multi-scale feature channels, and the low-level features are upsampled and concatenated with the high-level features before being input into the high-level network.

2. An online intelligent recommendation system for remote sensing algorithms in public interest litigation by procuratorates, characterized in that: It includes a user interface module, an offline processing module, a workflow library module, and an online recommendation module; The user interface module imports process data and provides modeling interaction; the offline processing module executes the steps described in claim 1; the process library module stores flowchart data, sequence data, and graph pair data; and the online recommendation module calls the trained recommendation model to generate node recommendation results.

3. The system according to claim 2, characterized in that, The user interface module configures process retrieval, editing, and version control to support incremental process updates and historical rollback.

4. The system according to claim 2, characterized in that, The online recommendation module uses a message queue to receive process prefixes and writes them into the process library after generating recommendation results for auditing and optimization.

5. A computer-readable storage medium storing instructions thereon, wherein a processor executes the instructions to perform the steps of the method of claim 1.

Images