Unattended processing system for multi-source remote sensing data

Abstract

The invention discloses an unattended processing system for multi-source remote sensing data, and relates to the technical field of remote sensing data processing. The system comprises an automatic production module, an intelligent management module, a cooperative control module and a three-stage storage pool, the automatic production module executes end-to-end large-scale processing from preprocessing to product generation on the multi-source remote sensing data based on a workflow template capable of being arranged, and supports parallel task level and data level; the intelligent management module realizes state awareness and self-adaptive resource scheduling through metadata driving and event monitoring; the cooperative control module serves as a center, provides a unified interface and manages configuration in a centralized mode. And the three-stage storage pool realizes automatic data circulation under the control of the system. Through event driving and intelligent scheduling, the whole-process unattended automatic processing of optical and radar remote sensing data from access to standard product output is realized, and the production efficiency, the system manageability and the expandability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing data processing technology, and more specifically, to an unattended processing system for multi-source remote sensing data. Background Technology

[0002] With the rapid development of high-resolution Earth observation technology, remote sensing data is growing at an unprecedented speed and scale. Efficiently transforming massive amounts of multi-source, heterogeneous raw satellite data into standard geographic information products that can be directly used in various industries has become a key infrastructure requirement for driving the development of the digital economy.

[0003] Currently, the industry generally adopts a semi-automated processing mode that heavily relies on manual operation and switching between multiple software programs. This mode requires technicians to manually import optical and radar data into separate systems, and to frequently perform manual configuration and intervention in core stages such as data preprocessing, project creation, quality control, and process integration. This traditional approach has systemic defects: its automation level is low, making it difficult to achieve large-scale continuous production; the management of intermediate data and parameters generated during data processing is chaotic, with poor traceability and difficulty in guaranteeing quality; data silos are formed between different processing systems, relying on manual bridging, resulting in loose processes that are prone to errors; and it lacks unified intelligent scheduling capabilities, making it impossible to dynamically allocate resources and respond to anomalies, resulting in insufficient overall efficiency and robustness.

[0004] Therefore, the industry urgently needs a brand-new solution that can completely break through the existing technological bottlenecks to achieve a fundamental transformation of multi-source remote sensing data processing from manual to fully automated and intelligent, and to meet the requirements of efficient, reliable, and manageable large-scale production. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, the present invention provides an unattended processing system for multi-source remote sensing data to address the problems mentioned in the background art.

[0006] In a first aspect, embodiments of this application provide an unattended processing system for multi-source remote sensing data, including: The system includes an automated production module, an intelligent management module, a collaborative control module, and a three-tiered storage pool connected through a data transfer process. The three-level storage pool includes the original storage pool, the intermediate storage pool, and the result storage pool; The original storage pool is equipped with a listening unit, which is communicatively connected to the message middleware. The intelligent management module includes a metadata database and a scheduler, and the scheduler is communicatively connected to the message middleware and the metadata database. The automated production module includes a workflow engine, a processing algorithm library, and a distributed computing node cluster. The workflow engine is communicatively connected to the scheduler and configuration center and is used to load processing flow templates and instantiate them into atomic task queues. The processing algorithm library encapsulates optical image processing routines and radar image processing routines. The collaborative control module includes a service gateway, a configuration center, and a status synchronization unit. The service gateway receives external requests, the configuration center stores the processing flow template, and the status synchronization unit is connected to the metadata database and the web frontend.

[0007] In some embodiments of this application, the workflow engine is the Apache Airflow engine; The distributed computing node cluster runs either the Apache Spark or Dask framework; The message middleware is either Redis or RabbitMQ; The processing algorithm library integrates the GDAL library, the OrfeoToolbox library, and the SNAP toolbox.

[0008] In some embodiments of this application, the processing flow template defines the dependencies and parallel relationships between data processing steps using a directed acyclic graph.

[0009] In some embodiments of this application, the metadata database is a PostgreSQL database with PostGIS extensions deployed; The monitoring unit monitors a specified directory of the original storage pool based on the Watchdog library.

[0010] In some embodiments of this application, the scheduler is configured with a first preset threshold and a second preset threshold; When the system's average CPU utilization rate is detected to be consistently higher than the first preset threshold or the length of the atomic task queue is detected to exceed the second preset threshold, the scheduler triggers a computing resource expansion operation.

[0011] In some embodiments of this application, the service gateway is a RESTful API gateway; The state synchronization unit establishes a persistent connection with the web frontend based on the WebSocket protocol.

[0012] In some embodiments of this application, the remote sensing data received by the original storage pool includes mainstream domestic and international optical satellite imagery data and radar satellite imagery data; The image products output by the results storage pool are digital orthophoto maps, digital surface models, or digital elevation models.

[0013] In some embodiments of this application, the radar image processing routine includes code modules for performing format conversion, multi-view processing, adaptive filtering, and RPC parameter generation.

[0014] In some embodiments of this application, the processing algorithm library also includes a smart mosaic line generation module; The intelligent mosaic line generation module constructs a mesh map from two input overlapping images, assigns a first weight to the mesh edges based on a semantic classification layer, assigns a second weight to the mesh edges based on pixel contrast, and uses a graph search algorithm to find the path from one side of the overlapping area to the other side with the minimum total weight as the mosaic line.

[0015] In some embodiments of this application, the configuration center supports the addition, deletion, and modification of the processing flow template; The processing algorithm library supports plug-in integration of new algorithm modules.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention completely changes the traditional serial operation mode of multiple manual interventions by constructing a fully automated processing pipeline based on event-driven and workflow orchestration. The system uses monitoring units deployed in the storage pool to automatically detect the arrival of data, and the scheduling core seamlessly triggers the entire chain from preprocessing to product generation, realizing unattended production.

[0017] 2. This invention innovatively designs a system architecture that deeply integrates "three-level storage and parallel scheduling." By tightly coupling logically separated raw, intermediate, and result storage pools with task scheduling services, it solves the pain points of chaotic data management and difficult process traceability in the traditional model. The intermediate storage pool, as a buffer and shared area for parallel computing, realizes efficient decoupling and data-level parallelism of upstream and downstream tasks in complex workflows.

[0018] 3. This invention proposes an intelligent mosaic line automatic generation method that integrates semantic and contrast information, overcoming the bottlenecks of traditional processes that rely on manual experience, are time-consuming, and produce inconsistent quality. Within a graph theory optimization framework, this method assigns a fusion cost based on semantic classification (e.g., high weight for buildings) and local pixel contrast to the image overlap area grid, and searches for the minimum cost path, accurately achieving automatic avoidance of important features and seamless visual transition of mosaic lines. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0020] Figure 1 The system data flow diagram provided by this invention; Figure 2 The processing flowchart provided by this invention; Figure 3 The three-level storage flow diagram provided by this invention; Figure 4 The flowchart of the intelligent mosaic line generation algorithm provided by the present invention. Detailed Implementation

[0021] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] Example 1 Please see Figure 1 This invention provides an overall architecture and data flow implementation scheme for an unattended processing system for multi-source remote sensing data. To achieve the technical objective of automatically, efficiently, and stably processing multi-source satellite data (optical and radar data) and ultimately producing standard orthophoto images and elevation models, this invention employs the following technical solution: The system mainly consists of a collaborative control module, an intelligent management module, an automated production module, and logically separated three-level storage pools. These modules are not simply software stacks, but are deeply coupled with the data flow through carefully designed interfaces, forming an organic whole characterized by event-driven, automatic scheduling, and pipelined operation. The system is deployed in a cluster environment composed of multiple high-performance servers. Each server is typically configured with an Intel 32-core or higher processor, at least 64GB of RAM, a 1TB solid-state drive for the system disk to ensure system service response speed, and 10-20TB hard disks for storing massive amounts of remote sensing data. The network environment must guarantee gigabit bandwidth to meet the needs of large-scale intermediate data exchange between cluster nodes.

[0024] Specifically, the collaborative control module serves as the sole external portal and command center for the entire system. Its core is a backend service cluster developed using frameworks such as Spring Boot or Django, with the frontend using Nginx for load balancing. This module exposes a unified RESTful API interface. Any user commands or external system calls, such as task submissions, status queries, and template management, must enter the system through this interface, achieving decoupling and security control between the frontend and backend. Internally, an independent configuration center service is responsible for the centralized and versioned management of all system operating parameters and the crucial "standardized processing flow templates." These templates are stored in JSON or YAML formats, defining complete processing chains and parameter sets for different satellite sensors. When adjustments to the processing flow or adaptation to new satellites are needed, only updating or adding templates in the configuration center is required, without modifying the code, greatly improving the system's maintainability and scalability. In addition, the module has an embedded state synchronization unit based on WebSocket technology. It continuously obtains real-time information such as global task status and node resource load from the metadata database of the intelligent management module and actively pushes it to the web front-end interface. This allows users to remotely and in real-time monitor the entire production cluster's operation in a browser, achieving transparent management "anytime, anywhere".

[0025] In this embodiment, the intelligent management module is the brain of the system, enabling "unattended operation" and "intelligent management." It interacts closely with the collaborative control module, receiving macro-level production instructions and transforming them into executable micro-level scheduling decisions. One of the core components of this module is the metadata database, built using PostgreSQL and spatially extended PostGIS. It records end-to-end, granular metadata information, including but not limited to: the source of raw data, sensor type, and imaging time; the creation time, execution status, resource consumption, associated algorithms and parameters of each processing task; the storage path, data quality indicators, and production logs of each piece of intermediate and final data. A unique task ID links all stages, parameters, and data entities throughout the data processing lifecycle, achieving complete traceability. Another core component of this module is the listening unit deployed on the raw storage pool server, developed based on file system event listening libraries such as Watchdog. It monitors the designated data ingestion directory 24 / 7. Once it detects that a new satellite raw data file has been copied and written, it immediately captures the "file creation" or "file closing" event, extracting basic information such as file path and size. Then, by parsing the filename or reading the accompanying XML metadata file, it obtains more precise information such as satellite identifiers and product levels, ultimately packaging it into a structured "data arrival event." This event is immediately published to message queues such as Redis or RabbitMQ. The scheduler service within the intelligent management module subscribes to this message queue, triggering subsequent automated processing pipelines upon receiving a new event. The scheduler also possesses dynamic resource scheduling capabilities, continuously collecting metrics such as CPU utilization, memory usage, and task queue length from monitoring systems like Prometheus. When it detects that the average CPU load consistently exceeds 80% (first preset threshold) or the backlog of pending tasks exceeds 50 (second preset threshold), the scheduler determines that current computing resources are insufficient and automatically scales up new computing nodes (virtual machines or containers) by calling cloud platform APIs or cluster management tools, automatically deploying the necessary data processing services on the new nodes. When a task fails, the scheduler will automatically retry 1-3 times according to a preset strategy. If it still fails, it will be escalated to an alarm, notifying management personnel to intervene, thereby achieving "self-healing" of the process and ensuring production continuity.

[0026] In this embodiment, the automated production module is responsible for performing specific, high-intensity data processing calculations. It is not a single software application, but rather a cluster of horizontally scalable "processing service instances." Each instance runs in a Docker container, encapsulating specific processing capabilities, such as "optical radiometric calibration service," "radar multi-view processing service," and "regional network adjustment service." These services communicate with the scheduler through standard interfaces and, when idle, "pull" their own atomic tasks from the global task queue. The core engine of the module is Apache Airflow, which is responsible for instantiating the "standardized processing flow templates" defined in the configuration center into specific, executable workflows (DAGs). For example, for a template of "producing orthorectified images from Gaofen-2 L1A level data," Airflow will parse it into a directed acyclic graph containing multiple nodes such as "data decompression," "radiometric calibration," "geometric coarse correction," "RPC generation," "regional network adjustment," "orthorectification," "color fusion," "smart mosaicking," and "cropping and segmentation." Each node is an atomic task. Airflow's scheduler comprehensively considers task dependencies and resource availability, distributing these node tasks to idle processing service instances. The actual computation relies on a powerful distributed computing framework and underlying algorithm libraries. For example, in orthorectification or large-scale mosaicking stages, Spark can divide a large image into blocks and distribute them across multiple cores in the cluster for parallel processing, achieving data-level parallelism. Simultaneously, multiple different production tasks are completely independent and can be executed in parallel by different service instances in the cluster, achieving task-level parallelism. The processing algorithm libraries integrated into this module mainly include GDAL, OrfeoToolbox, and SNAP. These open-source tools are encapsulated in a unified microservice interface and uniformly scheduled by the task scheduling core, thereby integrating heterogeneous algorithm capabilities into a coherent pipeline.

[0027] In this embodiment, the three-tiered storage pool acts as the "blood vessels" of the system, defining the flow trajectory and temporary storage location of data throughout the production lifecycle. It is a key architectural design supporting efficient parallel processing. The raw storage pool is a high-performance network attached storage (NAS) or object storage mount directory, specifically used to receive raw L0 / L1 level data pushed from satellite ground receiving stations or manually imported. The intermediate storage pool is also based on high-performance shared storage and is used to store all intermediate results generated during processing, such as radiometrically calibrated but uncorrected images, tie point files after regional network adjustment, and orthorectified single-scene images. Its existence is crucial: First, it decouples upstream and downstream processing tasks. Upstream tasks only need to write the results to the intermediate pool to mark them as complete, and downstream tasks can read the required data from the intermediate pool at any time without synchronous waiting, forming an asynchronous pipeline. Second, it serves as "shared memory" for parallel computing. When a task needs the same intermediate data, multiple computing nodes can directly read it from the intermediate pool, avoiding redundant computation and data copying. The results storage pool stores the final, user-compliant standard products, such as GeoTIFF format digital orthophoto maps (DOM) and digital terrain models (DSM). The three storage pools are connected via an automated data transfer process controlled by the task scheduling core. For example, when a "data arrival event" is triggered, the scheduler not only initiates the processing flow but also instructs the raw data to be moved from the receiving directory of the "raw storage pool" to its working directory. During processing, each atomic task's service instance is explicitly instructed to read input data from which storage pool and at which path, and to write output data to which storage pool and at which path. Data thus flows orderly and automatically across the three storage pools according to the logic defined in the processing flow template, until it becomes the final product.

[0028] Example 2 Please see Figure 2 This invention provides an implementation scheme for an end-to-end automated processing workflow from data to product in an unattended processing system for multi-source remote sensing data. To transform the traditionally complex remote sensing processing workflow, which relies on manual, step-by-step operations and frequent interventions, into a single-trigger, fully automated, parallel, and highly efficient uninterrupted pipeline, this invention achieves this through the following technical solution: The workflow consists of a series of standardized, programmable processing steps, automatically triggered and driven by an event monitoring and scheduling system. The entire process begins with the arrival of data and ends with the product entering the warehouse, requiring no manual operation nodes in between. However, it retains an entry point for interactive manual editing via the web interface of the collaborative control module when necessary, serving as a supplementary guarantee to the automated workflow.

[0029] Specifically, the process begins with automatic data import and event triggering. When the satellite ground station pushes newly received raw data to a designated directory in the system's raw storage pool via FTP or a dedicated link, the monitoring unit described in Example 1 immediately captures this event. The monitoring unit does not simply issue a notification; it performs preliminary intelligent identification: based on file naming conventions or internal metadata, it automatically determines that the data originates from the "Gaofen-2" (GF2) optical satellite and is of product level L1A. Subsequently, it generates a structured "data arrival event," which includes key metadata such as sensor type, product level, and data path, and publishes it to the message queue. The scheduler in the intelligent management module consumes this event, marking the official start of the automated process.

[0030] Next, the automatic template matching and instantiation phase begins. The scheduler parses the metadata in the event, especially the "sensor type" and "product level" fields. It then queries the configuration center to find a matching "standardized processing workflow template." The system has a rich library of pre-built templates, such as "GF2_L1A_to_DOM" (producing digital orthorectified imagery from Gaofen-2 L1A level data) and "GF3_SLC_to_DOM" (producing digital orthorectified imagery from Gaofen-3 SLC level radar data). Upon successful matching, the scheduler calls the Apache Airflow interface to instantiate the template into a specific workflow instance and assigns it a unique task ID. This workflow instance already contains a complete directed acyclic graph (DAG) definition from preprocessing to the final product.

[0031] Subsequently, the system enters the fully parallel execution phase. Airflow's workflow engine begins scheduling the various nodes in the DAG. The entire process exhibits a high degree of parallelism: Task initialization and data assignment: First, an initialization task is triggered, which dynamically divides the large production area into several "job blocks" that can be processed in parallel based on the processing scope and the list of available data.

[0032] Parallel preprocessing: Preprocessing tasks are performed in parallel for different sensors. For radar data (such as Gaofen-3), multiple processing service instances are launched in parallel to perform format conversion, multi-view processing, adaptive filtering, and generate RPC parameter files for subsequent geometric processing. For optical data, radiometric calibration and sensor correction are performed in parallel. These preprocessing tasks are independent of each other, making full use of cluster computing resources.

[0033] Project Creation and Regional Network Adjustment: After preprocessing of multiple images within a certain area, the system automatically and intelligently delineates the task area based on the geographical location and overlap of the images, and creates a regional network adjustment project. Subsequently, the system automatically matches a large number of connection points and control points throughout the entire area. The key innovation lies in the introduction of automated quality inspection and interactive editing interfaces. Steps that traditionally required manual point-by-point inspection and editing are now automatically completed by the system's built-in quality assessment algorithm. It automatically filters out and removes or adjusts gross errors based on indicators such as point residuals and correlation. Only in areas where the automatic quality inspection algorithm has low confidence will the system mark them in the task log, allowing production personnel to quickly locate and perform a small amount of interactive editing through the web interface, rather than participating in the entire process. The adjustment calculation itself is also a highly parallel process, utilizing multi-core servers or even distributed computing frameworks to solve large-scale equations.

[0034] Orthorectification, fusion, and mosaicking: After adjustment, a high-precision geometric model is obtained. Orthorectification is then performed on each image, a process that exemplifies data-level parallelism, allowing multiple images to be distributed to different computing nodes for simultaneous processing. For optical satellites, subsequent color balancing and sharpening fusion processing may be performed. One of the most valuable automation steps in the process is mosaicking. Traditionally, manual drawing of mosaic lines in overlapping areas is required to avoid buildings and ensure natural tone transitions, which is time-consuming and labor-intensive. This system employs an intelligent mosaic line generation algorithm (the principle of which is detailed in Example 4) to automatically find the optimal mosaicking path. The system automatically performs mosaicking preprocessing and intelligently finds and generates seamless mosaic lines based on image content, completely replacing manual drawing.

[0035] Cropping and Product Storage: After mosaicking is completed, the system automatically crops the large-scale mosaic results into regular sectional products based on standard map sizes or custom ranges. Finally, all final product files, quality reports, production logs, etc., are automatically transferred and registered in the results storage pool, and the metadata database is updated to "Completed". Users will receive a notification of task completion via the web interface and can immediately download the products.

[0036] In this embodiment, the entire workflow is driven entirely autonomously by the system. From data arrival to event triggering, template matching, and parallel execution, each step is automatically connected through inter-service API calls, message queues, and database status updates, forming a robust and efficient automated pipeline. The human role shifts from "operator" to "supervisor and configurer," requiring only the configuration of templates and rules during system initialization, or the occasional handling of edge cases that the automated process cannot resolve. This reduces manual operation time from 6-8 hours per day in the traditional model to approximately 0.5 hours, achieving truly "unattended" batch production.

[0037] Example 3 Please see Figure 3 This invention provides an unattended processing system for multi-source remote sensing data, supporting a three-tier storage architecture and a data collaborative flow implementation scheme for efficient parallel processing and pipelined operations. To achieve efficient, orderly, and low-latency flow of massive amounts of remote sensing data in complex processing chains, and to solve problems such as chaotic data management, loss of intermediate states, and I / O becoming a performance bottleneck in traditional models, this invention achieves this through the following technical solutions: constructing a logically clearly separated, physically configurable three-tier storage pool consisting of raw, intermediate, and finished data; and using a metadata-driven data transfer strategy to enable data to automatically and accurately flow between the three levels of storage under the drive of processing tasks.

[0038] Specifically, the raw storage pool primarily handles data reception and long-term archiving. Its storage medium can be a low-cost, high-capacity hard disk array, typically organized as a file system or object storage bucket. All raw satellite data input from external sources is first stored in a specific directory within this pool. The monitoring unit monitors this inbound directory. Data in the raw storage pool is typically read-only; processing tasks never directly modify the raw data, ensuring the integrity and reproducibility of the data source. Once the data has been successfully processed and archived, it can be migrated to cheaper offline storage based on storage strategies.

[0039] The intermediate storage pool is the innovative core and performance key of the entire system architecture, specifically designed to store all temporary data and intermediate results generated during processing. Its design needs to meet the requirements of high throughput and low latency concurrent read / write operations, therefore it typically employs a distributed file system accelerated by high-speed SSD caching or a high-performance NAS. The directory structure of the intermediate storage pool is carefully designed and strongly correlated with task IDs and processing stages. For example, a path might be organized as / intermediate / task_20231010001 / 1_preprocess / GF2_001.tif, clearly indicating the task, production stage, and specific content of the data. This design brings multiple benefits: First, it achieves task decoupling. For example, the preprocessing task completes once the corrected image is written to the intermediate pool, without waiting for and blocking the orthorectification task. The orthorectification task can retrieve the required data from the intermediate pool for calculation at any time. Second, it enables data sharing and reuse. A single adjustment parameter file generated by the "regional network adjustment" stage can be concurrently read by all orthorectification tasks participating in the adjustment, avoiding duplicate data generation and transmission. Furthermore, it optimizes parallel computing. During parallel data processing, the master node can easily distribute a list of 100 input file paths stored in the intermediate pool to 100 worker nodes. Each node reads, computes, and writes back independently, significantly reducing I / O contention. Finally, it enhances traceability and debugging capabilities. Any intermediate results are persistently stored. If a problem is found in the final product, the path in the intermediate pool can be traced back to check which stage's input or output was incorrect, allowing for precise problem localization.

[0040] The results storage pool is used to store the final delivered standard products. It has the highest requirements for data integrity, consistency, and long-term availability, but lower requirements for write frequency than the intermediate storage pool. It is typically managed using stable and reliable object storage or databases. Product entry into the database is not a simple file copy, but a "registration" process: in addition to writing the final DOM and DSM files to the results pool, the system also records detailed metadata of the product in the metadata database, including spatial range, production time, algorithm version used, quality assessment report, thumbnail preview path, etc., and associates it with the original data and processing task ID to form a complete data genealogy.

[0041] In this embodiment, the automatic data transfer between the three levels of storage is jointly completed by the task scheduling core and various processing service instances. The transfer rules are encoded in a "standardized processing flow template". When Airflow schedules an atomic task, it passes the input data path and the specified output data path required by the task as parameters to the processing service instance executing the task. Before starting the computation, the service instance first checks whether the input data is already cached locally. If not, it automatically reads it from the specified storage pool path. After the computation is completed, the result is written to the output path specified in the template. The initial data transfer from the original pool to the intermediate pool is usually completed by a dedicated "data preparation" task, which may perform preliminary operations such as data decompression and format conversion. The entire transfer process is fully automated, without the need for manual specification or modification of any file paths, fundamentally eliminating data loss or processing failures caused by manual configuration errors. This deeply integrated "three-level storage-parallel scheduling" architecture allows for a perfect match between data flow and business flow, ensuring both processing flexibility and scalability while maximizing the performance of distributed computing. It is the core infrastructure supporting the system to achieve processing cycles ranging from days to hours.

[0042] Example 4 Please see Figure 4 This invention provides a specific implementation scheme for an intelligent mosaicking line generation algorithm used in the automated mosaicking process of an unattended processing system for multi-source remote sensing data. To replace traditional manual visual interpretation and delineation of mosaicking lines, and to achieve fully automatic, high-precision mosaicking line finding that intelligently avoids important features, this invention achieves this through the following technical solution: an algorithm based on graph theory optimization and incorporating a dual guidance mechanism of image semantic information and visual contrast features to automatically find an optimal boundary path within the overlapping area of ​​the image to be mosaicked.

[0043] Specifically, the algorithm takes two geometrically registered images with overlapping regions as input. First, the algorithm constructs a graph theory model within the overlapping region. The overlapping region is discretized into a dense grid, with each intersection of the grid considered a node in the graph. Adjacent nodes are connected by edges. In this way, the entire overlapping region is transformed into a grid graph.

[0044] The core innovation of the algorithm lies in assigning a cost weight to each edge in the graph. The goal is to find a path from one boundary of the overlapping region to the other, such that the sum of the weights of all edges traversed by this path is minimized. This path is the optimal mosaic line. The weight calculation is not random or based on simple gray-level differences, but is dynamically and fusedly determined by the following two rules: Semantic classification guidance rules: Before algorithm execution or during parallel processing, the system uses a pre-trained deep learning semantic segmentation model to classify ground features in the input image, obtaining a probability map of each pixel belonging to categories such as "buildings," "vegetation," "water bodies," "roads," and "bare land." When calculating edge weights, the algorithm queries the semantic category of the area traversed by the edge in both images. The core rule is to prioritize mosaic lines passing through low-value, low-visual-sensitivity areas. Therefore, edges traversing densely populated areas of artificial features such as "buildings" are assigned extremely high weights (e.g., weight = 10), because cutting buildings would produce obvious visual defects and geometric errors; edges traversing homogeneous, non-cuttable natural areas such as "vegetation" and "water bodies" are assigned lower weights (e.g., weight = 1-3); and medium weights are assigned to "roads" and "bare land." This rule ensures that the algorithm actively avoids built-up areas such as towns and villages globally.

[0045] Pixel contrast guidance rule: Relying solely on semantics may be insufficient in certain scenarios (such as semantic segmentation errors or homogeneous building roofs). Therefore, the algorithm also introduces local pixel contrast gradients as a supplementary rule. The difference in pixel values ​​at Img_A and Img_B between the two nodes connected by an edge is calculated. If the difference is large, it indicates that the tones and textures of the two images at that location are highly inconsistent; passing through this location would easily result in noticeable seams after mosaicking. Therefore, the algorithm adds extra weight to edges in high-contrast regions. The weight increment can be a linear or non-linear function of the contrast. This rule ensures that even within homogeneous semantic regions, the algorithm tends to choose the point where the two images look most similar, achieving a seamless color transition.

[0046] In this embodiment, the weights of the two rules mentioned above are combined into a unified cost function. For example, the final weight Cost(e) of an edge e can be expressed as: Cost(e) = α * Semantic_Cost(e) + β * Gradient_Cost(e), where α and β are hyperparameters that adjust the importance of the two rules and can be determined through experimental optimization. The semantically guided cost graph is thus constructed.

[0047] Subsequently, the algorithm applies the classic graph search shortest path algorithm to search this weighted mesh graph. The starting point of the search is the set of all nodes on one boundary of the overlapping region, and the ending point is the set of nodes on the opposite boundary. The algorithm efficiently calculates the path with the minimum total cost among all possible paths from the starting set to the ending set. This path is the optimal mosaic line found by the system.

[0048] Finally, the system uses the found mosaic line to fuse the two images. Techniques such as feathering or optimal stitching are typically employed to smoothly transition along both sides of the mosaic line, generating a seamless mosaic result. The algorithm described in this embodiment transforms the traditionally time-consuming and subjective process of finding mosaic lines, which relies heavily on human experience, into a fully automated, repeatable, and objectively optimized computational process. It not only frees personnel from tedious labor but also, due to the use of more scientific optimization criteria, produces mosaic lines that are often more reasonable and accurate than those drawn manually, especially when dealing with complex scenes such as urban areas, significantly improving the geometric and visual quality of the final product.

[0049] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An unattended processing system for multi-source remote sensing data, characterized in that, include: The system includes an automated production module, an intelligent management module, a collaborative control module, and a three-tiered storage pool connected through a data transfer process. The three-level storage pool includes the original storage pool, the intermediate storage pool, and the result storage pool; The original storage pool is equipped with a listening unit, which is communicatively connected to the message middleware. The intelligent management module includes a metadata database and a scheduler, and the scheduler is communicatively connected to the message middleware and the metadata database. The automated production module includes a workflow engine, a processing algorithm library, and a distributed computing node cluster. The workflow engine is communicatively connected to the scheduler and configuration center and is used to load processing flow templates and instantiate them into atomic task queues. The processing algorithm library encapsulates optical image processing routines and radar image processing routines. The collaborative control module includes a service gateway, a configuration center, and a status synchronization unit. The service gateway receives external requests, the configuration center stores the processing flow template, and the status synchronization unit is connected to the metadata database and the web frontend.

2. The unattended processing system for multi-source remote sensing data according to claim 1, characterized in that, The workflow engine is the Apache Airflow engine; The distributed computing node cluster runs either the Apache Spark or Dask framework; The message middleware is either Redis or RabbitMQ; The processing algorithm library integrates the GDAL library, the OrfeoToolbox library, and the SNAP toolbox.

3. The unattended processing system for multi-source remote sensing data according to claim 2, characterized in that, The processing flow template defines the dependencies and parallel relationships between data processing steps using a directed acyclic graph.

4. The unattended processing system for multi-source remote sensing data according to claim 1, characterized in that, The metadata database is a PostgreSQL database with PostGIS extensions deployed. The monitoring unit monitors a specified directory of the original storage pool based on the Watchdog library.

5. The unattended processing system for multi-source remote sensing data according to claim 1, characterized in that, The scheduler is configured with a first preset threshold and a second preset threshold; When the system's average CPU utilization rate is detected to be consistently higher than the first preset threshold or the length of the atomic task queue is detected to exceed the second preset threshold, the scheduler triggers a computing resource expansion operation.

6. The unattended processing system for multi-source remote sensing data according to claim 1, characterized in that, The service gateway is a RESTful API gateway; The state synchronization unit establishes a persistent connection with the web frontend based on the WebSocket protocol.

7. The unattended processing system for multi-source remote sensing data according to claim 1, characterized in that, The remote sensing data received by the original storage pool includes mainstream domestic and international optical satellite imagery data and radar satellite imagery data; The image products output by the results storage pool are digital orthophoto maps, digital surface models, or digital elevation models.

8. The unattended processing system for multi-source remote sensing data according to claim 1, characterized in that, The radar image processing routine includes code modules for performing format conversion, multi-view processing, adaptive filtering, and RPC parameter generation.

9. The unattended processing system for multi-source remote sensing data according to claim 1, characterized in that, The processing algorithm library also includes an intelligent mosaic line generation module; The intelligent mosaic line generation module constructs a mesh map from two input overlapping images, assigns a first weight to the mesh edges based on a semantic classification layer, assigns a second weight to the mesh edges based on pixel contrast, and uses a graph search algorithm to find the path from one side of the overlapping area to the other side with the minimum total weight as the mosaic line.

10. The unattended processing system for multi-source remote sensing data according to claim 1, characterized in that, The configuration center supports adding, deleting, and modifying the processing flow templates; The processing algorithm library supports plug-in integration of new algorithm modules.

Images