A remote sensing image intelligent preprocessing method and system based on a localized multi-modal visual model

Abstract

The application discloses a kind of based on local multi-modal visual model's remote sensing image intelligent preprocessing method and system, it is related to remote sensing image processing technical field, method includes: obtaining the metadata of input remote sensing image and user task target, image attribute data is parsed, quality preliminary judgment is carried out in local computing environment by calling lightweight visual analysis model, and pre-processing execution chain is dynamically generated in combination with image attribute, quality preliminary judgment result and task target, corresponding pre-processing operation is executed, quality score is carried out in key processing node, and when the score does not satisfy condition, adjust execution chain, output standardization remote sensing image data product.Through the technical scheme of the application, pre-processing process can be adaptively determined according to different remote sensing images and application targets, and closed-loop optimization is realized through local quality feedback, to improve the accuracy, automation degree and data security of pre-processing.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a remote sensing image intelligent preprocessing method based on a localized multimodal visual model and a remote sensing image intelligent preprocessing system based on a localized multimodal visual model. Background Technology

[0002] Remote sensing imagery has wide applications in resource surveys, natural environment monitoring, disaster assessment, geological analysis, and land space management. Before being used for quantitative analysis, thematic mapping, or subsequent intelligent interpretation, raw remote sensing imagery typically undergoes preprocessing steps such as radiometric calibration, atmospheric correction, geometric correction, orthorectification, image fusion, cloud detection, image mosaicking, cropping, and noise suppression to obtain standardized remote sensing imagery data products with accurate spatial location, reliable radiometric information, and uniform format.

[0003] Current remote sensing image preprocessing typically relies on specialized software and human experience. Operators need to manually select processing steps, set processing parameters, and organize the execution sequence based on sensor type, imaging time, band information, projection parameters, image quality status, and specific mission objectives. For multi-source, multi-temporal, and multi-scale remote sensing data such as Landsat, Sentinel, Gaofen series, and UAV multispectral imagery, the metadata structure, band configuration, spatial resolution, and quality issues of different data sources vary significantly. If manual configuration is used, not only is the processing efficiency low and repeatability poor, but improper parameter selection can also easily lead to unstable output product quality.

[0004] To improve processing efficiency, existing technologies include batch processing scripts, automated workflow platforms, and containerized remote sensing processing tools. These solutions typically rely on preset workflows or user-pre-edited processing chains, which can reduce manual execution steps to some extent. However, fixed workflows are difficult to dynamically adapt to the specific image attributes, cloud cover, noise, missing bands, texture complexity, and other quality conditions of the input image, as well as the user's task objectives. When images exhibit anomalies such as thin clouds, banded noise, geometric distortion, or obvious mosaic seams, existing automated workflows often cannot adjust processing steps, execution order, or processing parameters in a timely manner, still requiring manual intervention for judgment and correction.

[0005] Furthermore, the application of artificial intelligence technology in remote sensing is mostly concentrated in downstream interpretation tasks such as classification, object detection, change detection, or semantic segmentation, with less application in the overall perception, decision-making, and quality feedback of the preprocessing process. Even when some models can be used for cloud removal, super-resolution, or quality inspection, they are usually used as single processing tools, lacking the ability to automatically generate a preprocessing execution chain based on metadata parsing, initial quality assessment, and task objectives. They also lack a closed-loop mechanism for scoring intermediate results at key processing nodes and adjusting the processing chain accordingly.

[0006] Meanwhile, some advanced models or automated services rely on cloud deployment. For remote sensing images involving land resources, geological surveys, natural resource management, or other sensitive scenarios, uploading raw images or intermediate processing results to the cloud may pose data security risks, and network transmission will also increase processing latency, which is not conducive to localized, batch, or near real-time processing needs. Summary of the Invention

[0007] To address the aforementioned issues, this invention provides an intelligent preprocessing method and system for remote sensing images based on a localized multimodal visual model. By parsing the metadata of the input remote sensing image in a local computing environment and invoking a lightweight visual analysis model for preliminary quality assessment, the preprocessing workflow can simultaneously utilize image attributes, quality status, and user task objectives as decision-making criteria. Furthermore, it dynamically retrieves, evaluates, sorts, and assembles the preprocessing execution chain from a pre-set set of preprocessing operations. At key processing nodes, it adjusts parameters, replaces operations, or inserts remedial steps based on the quality scores of intermediate results. This avoids the problem of traditional pre-set workflows being unable to adapt to multi-source heterogeneous images and abnormal quality states, improves the automation, adaptability, and closed-loop optimization capabilities of remote sensing image preprocessing, reduces reliance on manual configuration, and enhances the quality stability of standardized remote sensing image data products.

[0008] To achieve the above objectives, this invention provides an intelligent preprocessing method for remote sensing images based on a localized multimodal visual model, comprising: Obtain the metadata of the input remote sensing image and the user's task objective, and parse the image attribute data of the input remote sensing image based on the sensor parameter knowledge base; A lightweight visual analysis model is invoked in the local computing environment to perform a preliminary quality assessment of the input remote sensing image, and preliminary quality judgment data is obtained. Based on the image attribute data, the initial quality assessment data, and the user task objective, multiple remote sensing image preprocessing operations are dynamically retrieved, evaluated, sorted, and assembled from a preset set of remote sensing image preprocessing operations to generate a preprocessing execution chain for the input remote sensing image. The corresponding remote sensing image preprocessing operation is executed according to the preprocessing execution chain, and standardized data flow is performed between each remote sensing image preprocessing operation according to the preset data interface. Intermediate processing results are obtained at key processing nodes of the preprocessing execution chain, and the intermediate processing results are scored for quality. If the quality score does not meet the preset quality conditions, the preprocessing execution chain is adjusted according to the quality score. The adjustment includes at least one of processing parameter adjustment, processing operation replacement, or remedial processing step insertion. The output is a standardized remote sensing image data product processed by the preprocessing execution chain.

[0009] In the above technical solution, preferably, the image attribute data includes at least one of the following: data source, sensor type, imaging time, band information, and projection parameters; The initial quality assessment data includes at least one of cloud coverage information, noise information, missing stripe information, image quality overview information, or image texture information.

[0010] In the above technical solution, preferably, the step of dynamically retrieving, evaluating, sorting, and assembling multiple remote sensing image preprocessing operations from a preset set of remote sensing image preprocessing operations based on the image attribute data, the preliminary quality assessment data, and the user's task objective, to generate a preprocessing execution chain for the input remote sensing image, includes: Based on the image attribute data and the user task objective, determine the basic preprocessing steps from the preset expert rule base; Based on the image attribute data, the initial quality assessment data, and the user task objective, the suitability of candidate preprocessing operations in the remote sensing image preprocessing operation set is evaluated to obtain the necessity weight of each candidate preprocessing operation and the recommended initial processing parameter range. Based on the basic preprocessing steps, the necessity weights, and the recommended initial processing parameter range, the candidate preprocessing operations are sorted and assembled according to their suitability and necessity weights to generate a preprocessing execution chain that includes processing steps, preprocessing operation identifiers, execution order, and initial processing parameters.

[0011] In the above technical solution, preferably, the set of remote sensing image preprocessing operations includes multiple operations such as radiometric calibration, atmospheric correction, geometric correction, orthorectification, image fusion, cloud detection and processing, image mosaicking, image cropping, noise suppression, image registration, or image enhancement. The multiple remote sensing image preprocessing operations are determined based on the user's task objective, the image attribute data of the input remote sensing image, and the initial quality assessment data.

[0012] In the above technical solution, preferably, the preprocessing execution chain is described as work order data in a preset format that includes multiple processing nodes and node dependencies; The step of performing the corresponding remote sensing image preprocessing operation according to the preprocessing execution chain includes: Construct a directed acyclic graph based on the node dependencies; Perform a topological sort on the directed acyclic graph; Based on the topology sorting results, the remote sensing image preprocessing operations corresponding to multiple processing nodes are scheduled sequentially, and processing nodes that do not have data dependencies on each other are scheduled in parallel. The remote sensing image preprocessing operation is encapsulated in the form of containerized services. The containerized service corresponding to the processing node is called according to the work order data in the preset format, and remote sensing image data, processing parameters and processing status data are transferred between different containerized services through the preset data interface. Record the execution status of each processing node, and control the startup of subsequent processing nodes based on the execution status.

[0013] In the above technical solution, preferably, before calling the lightweight visual analysis model in the local computing environment, a local model deployment process is performed in advance, the local model deployment process including: The pre-trained model used for preliminary assessment of remote sensing image quality is subjected to at least one of the following processes: model quantization, operator fusion, or computational graph optimization, to obtain a locally deployed model for local inference. The inference backend, computational accuracy, and image segmentation strategy of the locally deployed model are determined based on the CPU, GPU, or memory resources of the local computing device. The input remote sensing images, intermediate processing results, and standardized remote sensing image data products are all processed in a local computing environment and are not sent to the cloud model service.

[0014] In the above technical solution, preferably, the key processing node includes at least one of atmospheric correction node, geometric correction node, cloud detection processing node, or mosaic processing node; The quality scoring of the intermediate processing results includes: The quality assessment sub-model is invoked to quantify and score at least one of the following indicators of the intermediate processing results: radiometric consistency, geometric correction residual, cloud processing integrity, or mosaic seam visibility. The quality assessment sub-model shares the underlying feature extraction network with the lightweight visual analysis model or the artificial intelligence model used to perform at least one remote sensing image preprocessing operation.

[0015] In the above technical solution, preferably, adjusting the preprocessing execution chain based on the quality score includes: When the quality score does not meet the preset quality conditions, at least one of the following adjustment methods will be executed: Adjust the processing parameters of the current processing step corresponding to the key processing node, and re-execute the current processing step; Replace the remote sensing image preprocessing operation used in the current processing step with a backup preprocessing operation selected from the preset set of remote sensing image preprocessing operations, and re-execute the current processing step. A remedial processing step is inserted in the preprocessing execution chain at a position following the current processing step.

[0016] In the above technical solution, preferably, the output of the standardized remote sensing image data product after processing by the preprocessing execution chain includes: The standardized remote sensing image data product is output as a geospatial data file containing georeferenced information. The standardized remote sensing image data product is a geospatial data product that has undergone physical correction and optimization, and does not contain land cover classification results, target detection results, or semantic segmentation results. Generate a processing metadata report associated with the standardized remote sensing image data product. The processing metadata report includes at least one of the following: preprocessing execution chain, processing steps, processing parameters, quality scores of key processing nodes, and execution status. The standardized remote sensing image data product and the processing metadata report are associated and archived.

[0017] This invention also proposes an intelligent preprocessing system for remote sensing images based on a localized multimodal visual model, comprising: The metadata intelligent perception and parsing module is used to acquire the metadata of the input remote sensing image and the user's task objectives, and to parse the image attribute data of the input remote sensing image based on the sensor parameter knowledge base. The local preliminary quality assessment module is used to call a lightweight visual analysis model in the local computing environment to perform a preliminary quality assessment of the input remote sensing image and obtain preliminary quality assessment data. The dynamic decision-making module is used to dynamically retrieve, evaluate, sort, and assemble multiple remote sensing image preprocessing operations from a preset set of remote sensing image preprocessing operations based on the image attribute data, the preliminary quality assessment data, and the user's task objective, and generate a preprocessing execution chain for the input remote sensing image. The pipeline execution module is used to execute the corresponding remote sensing image preprocessing operation according to the preprocessing execution chain, and to perform standardized data flow between each remote sensing image preprocessing operation according to a preset data interface. The quality feedback module is used to obtain intermediate processing results at key processing nodes of the preprocessing execution chain, score the intermediate processing results for quality, and adjust the preprocessing execution chain according to the quality score when the quality score does not meet the preset quality conditions. The adjustment includes at least one of processing parameter adjustment, processing operation replacement, or remedial processing step insertion. The local model deployment module is used to deploy the lightweight visual analysis model in a local computing environment, and to ensure that the input remote sensing images, intermediate processing results and standardized remote sensing image data products are all processed in the local computing environment and are not sent to the cloud model service. The output module is used to output standardized remote sensing image data products after being processed by the preprocessing execution chain.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By acquiring the metadata of the input remote sensing image and the user's task objectives, and by parsing the image attribute data based on the sensor parameter knowledge base, the system can identify the basic information of the input image, such as the data source, sensor type, imaging time, band information and projection parameters. This provides a data basis for subsequent preprocessing operations and parameter configuration, reducing the reliance on manual identification of image types and manual configuration of processing procedures.

[0019] (2) By calling a lightweight visual analysis model in the local computing environment to perform a preliminary quality assessment of the input remote sensing image, it is possible to obtain preliminary quality judgment data such as cloud cover, noise, missing stripes, image quality overview or texture status. The preprocessing process not only selects based on metadata, but also makes dynamic decisions based on the actual quality status of the image, thereby improving the adaptability to abnormal images such as thin clouds, noise, and missing stripes.

[0020] (3) By dynamically retrieving, evaluating, sorting and assembling the preprocessing execution chain from the preset remote sensing image preprocessing operation set according to image attribute data, preliminary quality judgment data and user task objectives, remote sensing images with different data sources, different quality states and different application objectives can generate differentiated processing flows, thereby overcoming the problem of poor generalization ability of traditional fixed preprocessing flows and improving the matching of preprocessing step selection, execution order and processing parameters.

[0021] (4) By executing the corresponding remote sensing image preprocessing operations according to the preprocessing execution chain and performing standardized data flow between each operation according to the preset data interface, the processing operations such as radiometric calibration, atmospheric correction, geometric correction, orthorectification, cloud detection, mosaicking, cropping, and noise suppression can be executed in a unified data format, thereby improving the modularity, automated execution efficiency and reusability of the remote sensing image preprocessing process.

[0022] (5) By obtaining intermediate processing results and scoring quality at key processing nodes in the preprocessing execution chain, and adjusting processing parameters, processing operations or remedial processing steps when the quality score does not meet the preset quality conditions, a quality feedback loop that adjusts the preprocessing execution chain from the execution results can be formed, reducing output quality fluctuations caused by improper single parameter configuration, abnormal image quality or fixed process defects.

[0023] (6) By outputting standardized remote sensing image data products after preprocessing execution chain, and associating them with metadata reports of the processing process, the final results have a unified format, geographic reference information and traceable processing records, thereby providing standardized input data with stable quality and traceable process for subsequent quantitative inversion, thematic analysis or intelligent interpretation. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall architecture of a remote sensing image intelligent preprocessing method and system based on a localized multimodal visual model, as disclosed in an embodiment of the present invention. Figure 2 This is a schematic diagram of dynamic decision-making in the preprocessing execution chain according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the execution process of a remote sensing image preprocessing pipeline disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the local deployment and inference optimization process of a lightweight visual analysis model disclosed in one embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] The present invention will now be described in further detail with reference to the accompanying drawings: like Figure 1 As shown, a remote sensing image intelligent preprocessing method based on a localized multimodal visual model according to the present invention includes: The system acquires metadata from input remote sensing images and user task objectives, and parses the image attribute data of the input remote sensing images based on a sensor parameter knowledge base. Input remote sensing images can be satellite remote sensing images, UAV multispectral images, stereo image pairs, or other multi-source remote sensing data. Image attribute metadata may include accompanying sensor information, imaging time, band descriptions, spatial references, and projection files. User task objectives can be surface temperature inversion, vegetation index mapping, 3D modeling, land cover classification preprocessing, or other tasks requiring standardized image products as input. The sensor parameter knowledge base can pre-include parameter specifications and processing standards for mainstream on-orbit sensors such as the Landsat series, Sentinel series, and Gaofen series, enabling automatic parsing of metadata files from different sources and in different formats without manual configuration by operators. This fundamentally solves the tedious problem of manually setting parameters for each sensor in traditional processing.

[0027] A lightweight visual analysis model is invoked in the local computing environment to perform a preliminary quality assessment of the input remote sensing image, obtaining initial quality judgment data. This lightweight visual analysis model, deployed on a local computing device, can quickly scan the input image to identify cloud cover, noise, missing bands, texture complexity, or overall image quality status. This step enables the preprocessing workflow to not only rely on metadata but also make judgments based on the actual visual content of the image. Furthermore, all inference processes are completed within the local computing environment, without transmitting any image data to external networks, thus ensuring the security of classified or sensitive remote sensing data.

[0028] Subsequently, based on image attribute data, preliminary quality assessment data, and user task objectives, multiple remote sensing image preprocessing operations are dynamically retrieved, evaluated, sorted, and assembled from a pre-defined set of remote sensing image preprocessing operations to generate a preprocessing execution chain for the input remote sensing image. This preprocessing execution chain is not a fixed template but is dynamically formed based on the current image attributes, quality status, and task objectives. Different sensor types require fundamentally different processing steps, and different task objectives (such as land surface temperature inversion, land classification, and 3D reconstruction) require varying preprocessing depths and algorithm selections. Dynamically generating the execution chain ensures the processing scheme's high adaptability to multi-source, multi-temporal, and multi-scale images, addressing the core issues of rigid workflows and poor generalization capabilities in traditional automated systems. For example, for thermal infrared land surface temperature inversion tasks, an execution chain including radiometric calibration, atmospheric correction, cloud detection, and study area cropping can be prioritized; for UAV multispectral vegetation monitoring tasks, an execution chain including registration, orthorectification, mosaicking, and preprocessing of index products can be generated.

[0029] After generating the preprocessing execution chain, the corresponding remote sensing image preprocessing operations are executed according to the chain, and standardized data flow is performed between each preprocessing operation using a preset data interface. The input, output, parameters, and status data of each preprocessing operation are transmitted through a unified interface, enabling multiple heterogeneous processing operations to be executed seamlessly. The standardized data interface design ensures a consistent data exchange format between different processing operations, avoiding processing interruptions or data quality loss due to format incompatibility.

[0030] Intermediate processing results are acquired at key processing nodes in the preprocessing execution chain. These intermediate results are then scored for quality. If the quality score fails to meet preset quality conditions, the preprocessing execution chain is adjusted based on the score. Adjustments may include at least one of the following: adjusting processing parameters, replacing processing operations, or inserting remedial processing steps. For example, if insufficient cloud mask integrity is found after cloud detection, the cloud detection threshold can be readjusted or a more refined cloud repair step can be inserted. If the residual after geometric correction is too large, the registration algorithm can be changed or the control point parameters can be adjusted. This quality feedback loop upgrades the traditional post-verification model to in-process online monitoring and correction, effectively avoiding cascading errors caused by low-quality intermediate results being passed to subsequent steps. This significantly improves the robustness and consistency of the final output product when facing multi-source heterogeneous images or anomalous data.

[0031] The final output is a standardized remote sensing image data product after preprocessing execution chain.

[0032] In this implementation, by combining metadata parsing, local quality assessment, dynamic execution chain generation, standardized pipeline execution, and key node quality feedback adjustment, the remote sensing image preprocessing workflow can automatically adapt to the image's own state and task objectives. This reduces the reliance on manual configuration of processing workflows and parameters, overcomes the problem that fixed workflows are difficult to handle multi-source heterogeneous remote sensing data and abnormal quality images, and improves the automation, adaptability, and output quality stability of the preprocessing workflow.

[0033] In the above embodiments, preferably, the image attribute data includes at least one of the following: data source, sensor type, imaging time, band information, and projection parameters; the data source can be used to distinguish between Landsat, Sentinel, Gaofen series, UAV multispectral, or other sources; the sensor type can be used to determine the band combination, radiometric response characteristics, and geometric model; the imaging time can be used to assist in determining the solar altitude angle, seasonal changes, or temporal relationships, thereby assisting the atmospheric correction model in estimating the optical thickness of aerosols; the band information can be used to determine whether multispectral fusion, thermal infrared processing, or vegetation index-related preprocessing is required; the projection parameters can be used for subsequent geometric correction, orthorectification, and mosaic clipping to ensure that the final product has accurate georeference.

[0034] Preliminary quality assessment data includes at least one of the following: cloud cover information, noise information, missing stripe information, image quality overview information, or image texture information. Cloud cover information can be used to determine whether cloud detection, cloud masking, or cloud inpainting needs to be included in the preprocessing execution chain; noise information can be used to determine whether noise suppression operations need to be inserted; missing stripe information can be used to determine whether stripe inpainting is needed; image texture information can be used to assist in judging the difficulty of registration, matching, or 3D modeling processing, and it directly affects the selection of dense matching algorithms. Regions with poor texture require the introduction of more constrained matching algorithms or auxiliary data.

[0035] During implementation, image attribute data is used to address the questions of "what type of data is this and what basic preprocessing is suitable," while initial quality assessment data is used to address the questions of "whether there are any anomalies in the current image quality status and whether additional remedial processing is needed." Together, they provide the input basis for generating the dynamic preprocessing execution chain.

[0036] In this embodiment, by distinguishing and combining image attribute data and preliminary quality assessment data, preprocessing decisions can simultaneously consider the sensor attributes and actual quality status of remote sensing images, thereby improving the specificity of processing step selection and parameter recommendations and reducing process rigidity caused by relying solely on sensor type.

[0037] like Figure 2 As shown, in the above embodiment, preferably, when generating the preprocessing execution chain, the basic preprocessing steps are first determined from a preset expert rule base based on the image attribute data and the user's task objective. The preset expert rule base can record the correspondence between "sensor type - task objective - basic processing steps," solidifying the processing experience accumulated by domain experts over a long period into machine-executable knowledge. For example, if the sensor is a Landsat series and the user's task objective is surface temperature inversion, the basic preprocessing steps may include radiometric calibration and atmospheric correction; if the input is UAV multispectral imagery and the task objective is vegetation index mapping, the basic preprocessing steps may include registration, orthorectification, and mosaicking. The expert rule base provides a fast and reliable basic process framework, ensuring the compliance of the decision-making process in terms of professional logic and avoiding product quality defects caused by the omission of key necessary steps.

[0038] After determining the basic preprocessing steps, based on image attribute data, preliminary quality assessment data, and user task objectives, the suitability of candidate preprocessing operations in the remote sensing image preprocessing operation set is evaluated to obtain the necessity weight of each candidate preprocessing operation and the recommended initial processing parameter range. For example, when the preliminary quality assessment data indicates the presence of thin clouds, the necessity weight of cloud detection and cloud inpainting is increased; when the preliminary quality assessment data indicates the presence of stripe noise, the necessity weight of noise suppression or stripe inpainting operations is increased; when the task objective requires high geometric accuracy, the suitability of geometric correction and image registration operations is increased.

[0039] Subsequently, based on the basic preprocessing steps, necessity weights, and recommended initial processing parameter ranges, candidate preprocessing operations are sorted and assembled to generate a preprocessing execution chain containing processing steps, preprocessing operation identifiers, execution order, and initial processing parameters. This preprocessing execution chain includes both mandatory basic processing steps and enhanced or remedial processing steps dynamically added based on quality status and task objectives.

[0040] In this implementation, the expert rule base ensures that the preprocessing process has a professional foundation, while the adaptability assessment and necessity weight enable the process to change dynamically according to the specific image quality and task objectives, forming a non-preset execution chain that is different from a fixed template, thereby improving the adaptability of the preprocessing process to remote sensing images of multiple sources and multiple quality states.

[0041] In the above embodiments, preferably, the set of remote sensing image preprocessing operations includes multiple operations such as radiometric calibration, atmospheric correction, geometric correction, orthorectification, image fusion, cloud detection and processing, image mosaicking, image cropping, noise suppression, image registration, or image enhancement; the above operations can be used as standardized, pluggable processing units for dynamic decision-making processes.

[0042] Specifically, radiometric calibration converts the sensor's raw digital quantization values ​​into physically meaningful radiance values, serving as the prerequisite for subsequent atmospheric correction and all quantitative applications. Atmospheric correction eliminates the effects of atmospheric scattering and absorption on images, converting radiance values ​​into surface reflectance, which is crucial for quantitative applications such as land cover classification and change detection. Geometric correction and orthorectification eliminate geometric distortions caused by sensor attitude errors and terrain undulations, respectively, ensuring accurate geographic coordinates for the images. Image fusion combines the high spatial resolution of the panchromatic band with the rich spectral information of the multispectral band, enhancing the overall information content of the images. Cloud detection processing automatically identifies and masks cloud-covered areas, eliminating cloud contamination for subsequent analysis. Image mosaicking seamlessly stitches multiple images into a complete regional coverage data product. Image cropping extracts the required spatial range according to the boundaries of the study area. Noise suppression eliminates sensor noise or stripe noise caused by sensor malfunctions. Image registration achieves precise spatial alignment of multi-temporal or multi-source images. Image enhancement improves visual effects and highlights target features through contrast stretching, sharpening, and other techniques.

[0043] During implementation, multiple remote sensing image preprocessing operations are determined based on the user's task objectives and the image attribute data and preliminary quality assessment data of the input remote sensing images, rather than adopting a fixed full-scale execution strategy. For example, image fusion can be selected for images containing panchromatic and multispectral bands; orthorectification can be selected for images with complex terrain effects; image mosaicking can be selected for multi-scene stitching tasks; noise suppression can be selected for images with noise or striping issues; and image cropping can be selected for tasks that require outputting study area results.

[0044] This set of preprocessing operations is not required to be executed all at once, but rather serves as a library of candidate operations that can be dynamically retrieved and combined. The combination, order, and parameters of operations can differ for different input images and different task objectives.

[0045] In this implementation, by constructing an operation set that includes multiple types of remote sensing preprocessing operations, the dynamic decision-making module has a selectable and scalable processing foundation, which can support differentiated preprocessing in multi-source and multi-task scenarios, and improve the system's flexibility and scalability in processing complex remote sensing images.

[0046] like Figure 3 As shown, in the above embodiment, preferably, the preprocessing execution chain is described as work order data in a preset format containing multiple processing nodes and node dependencies; the work order data can adopt a structured format such as JSON to record processing nodes, node inputs and outputs, processing parameters, execution order and dependencies, so that the processing scheme generated by the dynamic decision engine can be transmitted to the execution module in a standardized form, ensuring lossless information transmission from decision to execution.

[0047] When performing corresponding remote sensing image preprocessing operations according to the preprocessing execution chain, a directed acyclic graph (DAG) is constructed based on node dependencies, and the DAG is then topologically sorted. Each node in the DAG corresponds to a specific preprocessing operation, and the directed edges between nodes express the data flow direction and sequential dependencies, logically defining the topological structure of the entire execution chain. This allows the scheduling engine to understand the order and parallel relationships between steps in a machine-readable manner. Topological sorting determines which processing nodes must be executed first and which can be executed after the preliminary data preparation is complete. For processing nodes that do not have data dependencies on each other, parallel scheduling can be performed. For example, if independent noise suppression operations in certain bands do not depend on the processing results of other bands, they can be executed in parallel; while atmospheric correction usually needs to be performed after radiometric calibration. Parallel scheduling fully utilizes the multi-core CPU and GPU resources of the local computing device, significantly reducing the overall processing time, especially for computationally intensive operations (such as atmospheric correction and dense matching).

[0048] Remote sensing image preprocessing operations are encapsulated as containerized services. Based on pre-formatted work order data, the system calls the corresponding containerized service for each processing node and transfers remote sensing image data, processing parameters, and processing status data between different containerized services through a pre-defined data interface. The execution status of each processing node is recorded and used to control whether subsequent processing nodes are started. Containerization completely packages the runtime environment, dependencies, and configurations required for each preprocessing operation, ensuring that different operations do not interfere with each other and are independently reusable.

[0049] Throughout the entire process, the execution status of each processing node is recorded in real time, and the startup of subsequent processing nodes is controlled based on the execution status. This ensures that subsequent operations are only triggered after all their dependent nodes have been successfully completed, guaranteeing the orderliness and reliability of the overall process and providing a node-level execution status information foundation for the quality feedback module.

[0050] In this implementation, the preprocessing execution chain is automatically executed through work order data, directed acyclic graphs, and containerized service scheduling. This improves the modularity, parallelism, and traceability of the preprocessing process, reduces integration difficulties caused by inconsistent environments for different algorithms, and improves batch processing efficiency.

[0051] like Figure 4 As shown, in the above embodiment, preferably, before calling the lightweight visual analysis model in the local computing environment, a local model deployment process is performed in advance. The local model deployment process includes: The pre-trained model used for preliminary assessment of remote sensing image quality is subjected to at least one of the following processes: model quantization, operator fusion, or computational graph optimization, to obtain a locally deployed model for local inference. Specifically, model quantization converts the parameters and activation values ​​of the pre-trained floating-point model from 32-bit floating-point representation to low-bit-width integer representation. This is done to calibrate the dataset and determine the dynamic range of activation values ​​for each layer. This significantly reduces the model's storage footprint and inference computation with minimal loss of accuracy, enabling visual analysis models that originally relied on high-end cloud servers to perform efficient inference on local ordinary computing devices. Operator fusion combines consecutive convolutional layers, normalization layers, and activation function layers in the computation graph into a single composite operator, reducing the number of memory read / write operations for intermediate results and thus reducing inference latency. Computation graph optimization can reduce redundant computations and invalid nodes through graph-level optimization techniques such as dead code elimination and constant folding, thereby improving inference efficiency.

[0052] Based on the CPU, GPU, or memory resources of the local computing device, the inference backend, computational precision, and image partitioning strategy for the locally deployed model are determined. For example, when the local device is equipped with a GPU with sufficient video memory, the GPU inference backend is preferentially selected to obtain the optimal computational throughput. When only the CPU is available, the CPU inference backend is switched to and combined with thread-level parallel optimization. For large-format remote sensing images whose size exceeds the video memory capacity for a single inference, a partitioning strategy is adopted to divide the image into several overlapping sub-blocks, infer them in batches, and then merge the results, ensuring that any image size can be processed within limited video memory. This adaptive resource scheduling mechanism enables the method to remain available under different hardware configurations, from workstation-level servers to portable mobile terminals, giving the system extremely high deployment flexibility.

[0053] In this implementation, the input remote sensing images, intermediate processing results, and standardized remote sensing image data products are all processed in the local computing environment and are not sent to the cloud model service. This approach meets the stringent requirement of high-security scenarios that data must remain locally, and is particularly suitable for scenarios with high requirements for remote sensing data security and local processing timeliness, such as natural resources, land and space surveys, geological surveys, and disaster monitoring.

[0054] In this implementation, by model quantization, operator fusion, computation graph optimization, and local resource adaptation, the lightweight visual analysis model can complete inference in a local environment, which can reduce reliance on cloud transmission, reduce the risk of data leakage and network latency, and improve local batch processing and near real-time preprocessing capabilities.

[0055] In the above embodiments, preferably, the key processing nodes include at least one of atmospheric correction nodes, geometric correction nodes, cloud detection processing nodes, or mosaic processing nodes. The selection of these nodes is based on domain knowledge of remote sensing image processing, deploying them at key locations where the risk of error accumulation in the process is highest and subsequent steps are most sensitive to the quality of their results. Taking the atmospheric correction node as an example, the quality of atmospheric correction directly determines the accuracy of subsequent surface reflectance products. If there is a deviation here, all subsequent quantitative analyses based on reflectance will be systematically affected. The quality of the geometric correction node directly affects the overlay accuracy between the image and the reference coordinate system, thus affecting the registration, fusion, and measurement results of multi-source images. The quality of the cloud detection processing node determines the degree of cloud contamination removal in the image; residual cloud traces will introduce false signals in downstream classification or detection tasks. The quality of the mosaic processing node is reflected in the radiometric consistency and geometric continuity of the multi-image stitching area. Performing quality assessment at key nodes, rather than all nodes, ensures effective monitoring of key errors while avoiding excessive quality inspection overhead in local environments with limited computing resources.

[0056] When scoring the quality of intermediate processing results, a quality assessment sub-model is invoked to quantify at least one of the following indicators: radiometric consistency, geometric correction residuals, cloud processing integrity, or mosaic seam visibility. For example, atmospheric correction can evaluate whether the surface reflectance or brightness temperature value is within a reasonable range; geometric correction can evaluate control point residuals or registration errors; cloud detection processing can evaluate cloud mask integrity; and mosaic processing can evaluate seam visibility or color consistency.

[0057] The quality assessment sub-model shares a low-level feature extraction network with the lightweight visual analysis model or the AI ​​model used to perform at least one remote sensing image preprocessing operation. Sharing the low-level feature extraction network means that the quality assessment sub-model does not need to extract features from intermediate processing results from scratch; instead, it reuses the low-level visual feature representations already learned by the main processing model, only needing to add a lightweight head network for quality scoring. By sharing the low-level feature extraction network, redundant feature computation is reduced, resulting in lower computational overhead for quality scoring and maintaining consistency in feature representation with the main processing model.

[0058] In this implementation, by setting up a quality assessment sub-model at key processing nodes, the preprocessing process can identify intermediate result quality problems in a timely manner during execution, thereby improving the timeliness of quality anomaly detection and providing a quantitative basis for subsequent parameter adjustments, operation replacements, or remedial steps.

[0059] In the above embodiments, preferably, when adjusting the preprocessing execution chain according to the quality score, if the quality score does not meet the preset quality conditions, at least one adjustment method is executed.

[0060] The first adjustment method involves adjusting the processing parameters of the current processing step corresponding to the key processing node and re-executing the current processing step. Parameter adjustment is the least expensive correction method and is suitable for situations where the current algorithm itself has sufficient processing capacity but the parameters are inappropriately selected. For example, if the cloud detection mask is incomplete, the cloud detection threshold or window parameters can be adjusted and the cloud detection can be re-executed; if the geometric correction residual is large, the control point filtering parameters or matching threshold can be adjusted and the process can be re-executed.

[0061] The second adjustment method is to replace the remote sensing image preprocessing operation used in the current processing step with a backup preprocessing operation selected from the preset remote sensing image preprocessing operation set, and then re-execute the current processing step. Backup algorithm switching is applicable when the current algorithm has fundamental limitations in this type of image or under the current abnormal mode, and parameter adjustments cannot improve the results. For example, if a certain de-clouding algorithm is insufficient for thin cloud processing, and the cloud processing integrity score of the node after cloud detection processing reveals a large number of missed thin clouds, the original cloud detection algorithm based on radiation thresholds is no longer suitable for the cloud morphology and thin cloud spectral characteristics of the current image. In this case, the system automatically replaces it with a deep learning-based thin cloud detection model from the preprocessing operation set, utilizing its stronger feature extraction capabilities to complete thin cloud identification without manual intervention to judge and manually switch processing tools.

[0062] The third adjustment method is to insert a remedial processing step in the preprocessing execution chain, following the current processing step. Remedial step insertion is suitable when the processing result of the current step has not reached its optimal state but cannot be further improved, and needs to be compensated for in subsequent processes. For example, if cloud effects are still found after cloud detection, a cloud shadow repair or local enhancement step can be inserted; if obvious seams are found after mosaicking, a color balancing or seam optimization step can be inserted to perform local radiometric normalization on the repaired area, minimizing the impact of seams and repair marks on downstream analysis.

[0063] In this implementation, the preprocessing execution chain is not just executed once, but can be adjusted online based on the quality scores of key nodes, forming a closed loop of quality feedback in the processing process. This improves the system's robustness in the face of abnormal images, complex terrain, and diverse quality issues, and reduces manual rework and output quality fluctuations.

[0064] In the above embodiments, preferably, when outputting the standardized remote sensing image data product after preprocessing by the execution chain, the standardized remote sensing image data product is output as a geospatial data file containing georeferenced information. This geospatial data file can be in formats such as GeoTIFF, and retains information such as spatial reference, projection, resolution, and geographic extent. GeoTIFF embeds complete georeferenced information such as coordinate systems and geographic transformation matrices on top of the standard TIFF image format, enabling downstream quantitative inversion software, geographic information system tools, or deep learning inference platforms to directly read and use the product without additional format conversion or coordinate registration operations, significantly lowering the barrier to use after product delivery.

[0065] Standardized remote sensing image data products are geospatial data products that have undergone physical correction and optimization. They only include physical processing results such as radiometric calibration, atmospheric correction, and geometric correction, and do not include land cover classification results, target detection results, or semantic segmentation results. In other words, the output of this method focuses on preprocessed results for subsequent quantitative inversion, thematic analysis, or intelligent interpretation, rather than directly outputting downstream interpretation conclusions.

[0066] Simultaneously, a processing metadata report associated with the standardized remote sensing image data product is generated. This report includes at least one of the following: preprocessing execution chain, processing steps, processing parameters, quality scores for key processing nodes, and execution status. Subsequently, the standardized remote sensing image data product and the processing metadata report are archived together. This linked archiving ensures that product files and their quality reports are always stored in pairs, avoiding management chaos such as mismatched products and reports, and untraceable data sources in mass production scenarios. It provides a standardized data management foundation for subsequent data management, quality auditing, results sharing, and long-term data reuse.

[0067] In this implementation, by outputting geospatial data files and processing metadata reports, the preprocessing results possess dual attributes of data products and process records, which can improve the standardization, traceability, and reusability of the results, and provide input data with stable quality and transparent processing for subsequent remote sensing analysis tasks.

[0068] like Figure 1 As shown, this invention also proposes an intelligent preprocessing system for remote sensing images based on a localized multimodal visual model, comprising: The metadata intelligent perception and parsing module is used to acquire metadata of the input remote sensing image and user task objectives, and to parse the image attribute data of the input remote sensing image based on the sensor parameter knowledge base. This module is equivalent to the data understanding entry point of the system, and is used to identify the data source, sensor type, imaging time, band information and projection parameters of the input image.

[0069] The local preliminary quality assessment module is used to call a lightweight visual analysis model in the local computing environment to perform a preliminary quality assessment of the input remote sensing image and obtain preliminary quality assessment data. The dynamic decision-making module is used to dynamically retrieve, evaluate, sort, and assemble multiple remote sensing image preprocessing operations from a preset set of remote sensing image preprocessing operations based on image attribute data, preliminary quality assessment data, and user task objectives, and generate a preprocessing execution chain for the input remote sensing image. The pipeline execution module is used to perform corresponding remote sensing image preprocessing operations according to the preprocessing execution chain, and to perform standardized data flow between each remote sensing image preprocessing operation according to the preset data interface; The quality feedback module is used to obtain intermediate processing results at key processing nodes of the preprocessing execution chain, score the intermediate processing results for quality, and adjust the preprocessing execution chain according to the quality score when the quality score does not meet the preset quality conditions. The adjustment includes at least one of the following: adjusting processing parameters, replacing processing operations, or inserting remedial processing steps. The local model deployment module is used to deploy lightweight visual analysis models in a local computing environment, and to process input remote sensing images, intermediate processing results and standardized remote sensing image data products in the local computing environment without sending them to the cloud model service. The output module is used to output standardized remote sensing image data products after preprocessing execution chain.

[0070] In this implementation, the system modules sequentially form a collaborative structure of perception analysis, local initial quality assessment, dynamic decision-making, pipeline execution, quality feedback, local deployment, and output. The modules work collaboratively under a unified closed-loop architecture, enabling the automatic generation, execution, and quality closed-loop adjustment of the remote sensing image preprocessing workflow in a local environment, thereby improving the security, intelligence, and output product stability of multi-source heterogeneous remote sensing image preprocessing.

[0071] The intelligent preprocessing system for remote sensing images based on a localized multimodal visual model disclosed in the above embodiments has modules whose functions correspond to the steps of the intelligent preprocessing method for remote sensing images based on a localized multimodal visual model disclosed in the above embodiments. In the implementation process, the above embodiments are referred to for operation, and will not be repeated here.

[0072] The intelligent preprocessing method and system for remote sensing images based on a localized multimodal visual model disclosed in the above embodiments will be described in detail through the following examples.

[0073] Example 1: Processing Landsat 9 imagery for urban heat island effect analysis Input: The user uploads a Landsat 9 L1 level dataset and specifies the processing target as "for surface temperature inversion and urban heat island analysis".

[0074] Intelligent Sensing and Analysis: The metadata analysis module identified the sensor as "Landsat 9 OLI / TIRS". Visual analysis initially detected moderate thin clouds in the image and identified typical areas of the city and water bodies.

[0075] Intelligent dynamic decision-making: The decision engine integrates rules (thermal infrared data requires radiometric calibration and atmospheric correction) and model evaluation (cloud coverage, target is temperature inversion) to generate a customized workflow: [Radiometric calibration (including thermal infrared band)] → [Cloud detection and labeling (focusing on thin clouds)] → [Atmospheric correction (for thermal infrared, single window algorithm parameter estimation)] → [Study area clipping based on vector or interest points].

[0076] Automated execution: The pipeline engine sequentially calls the corresponding containerized services to perform calibration, cloud detection, atmospheric correction, and pruning. The computationally intensive atmospheric correction task is accelerated by GPUs.

[0077] Quality Assessment and Feedback: The quality self-inspection module assesses mask integrity after cloud detection and checks the reasonable range of brightness temperature values ​​for the land portion (e.g., -10℃ to 50℃) after atmospheric correction. If an abnormally high brightness temperature value is found in a certain area (possibly due to cloud detection omissions), a feedback signal is sent, and the decision engine may instruct a more refined cloud detection and correction process to be repeated for that area.

[0078] Output: The system outputs fully preprocessed GeoTIFF products of surface reflectivity and brightness temperature in each band, which can be directly used for surface temperature inversion, along with a PDF processing report containing process records, key parameters, and quality assessment scores.

[0079] Example 2: Batch processing of UAV multispectral imagery for precision agriculture Input: The user uploads multiple RAW format multispectral images and POS data from a folder.

[0080] Intelligent perception and analysis: The system identifies it as "five-band UAV multispectral image", visual analysis detects obvious lens edge distortion, and determines that the image overlap meets the stitching requirements.

[0081] Intelligent dynamic decision-making: For the goal of "precision agriculture (NDVI required)", the decision engine plans the workflow as follows: [lens distortion correction] → [multispectral band registration] → [structure from motion (SFM) and orthomosaic] → [radiative normalization (eliminating light differences between different frames)] → [calculate NDVI (normalized vegetation index)] → [clipping according to the imported field vector boundary].

[0082] Automated execution and evaluation: The pipeline engine calls upon computer vision libraries for correction and registration, and employs the SFM algorithm for sparse point cloud reconstruction, dense point cloud generation, and orthophoto mosaicking. The quality module evaluates the smoothness of the seams and the reasonableness of the NDVI values ​​at the boundaries of adjacent fields. If a seam is detected, it is fed back to the decision engine, which may add a mild color-balancing filter step.

[0083] Output: The system outputs a standard NDVI distribution map (GeoTIFF) for each specified field, as well as an NDVI orthophoto mosaic of the entire region, to guide variable fertilization or irrigation.

[0084] Example 3: Processing Gaofen-7 stereo image pairs for 3D modeling Input: The user uploads the front-view and rear-view panchromatic stereo image pairs from Gaofen-7 and the corresponding RPC files.

[0085] Intelligent perception and analysis: The system identifies the image as a "high-resolution stereo image pair", extracts the initial RPC parameters, and performs preliminary visual analysis to assess the texture richness of the image.

[0086] Intelligent Dynamic Decision Making: For the goal of "high-precision 3D reconstruction," the decision engine generates the core workflow: [Image-to-radioconsistency adjustment] → [High-precision dense matching to generate a disparity map] → [Generate a digital surface model (DSM) using RPC and disparity] → [DSM filtering (removing buildings and vegetation) to obtain a digital elevation model (DEM)]. Based on the texture richness evaluation results, the decision engine selects either a "semi-global matching (SGM)" or a "deep learning-based matching" model for the dense matching step.

[0087] Automated execution and evaluation: The engine calls the matching service and DSM generation service. The quality evaluation module checks the continuity of the disparity map and the effect of occlusion area processing. If a large area of ​​weak texture is found to be unmatched, feedback is given to the decision engine, which may try to switch the matching algorithm or introduce auxiliary data (such as an existing low-precision DEM) to constrain the area.

[0088] Output: The system outputs a high-precision DEM product (GeoTIFF) with a matching confidence plot.

[0089] Example 4: Handling Abnormal Sentinel-2 Data (Demonstrating Adaptive Advantages) Input: A user uploads a scene of Sentinel-2 L1C data for land cover classification, but the data also contains thin clouds and strip noise in some bands.

[0090] Intelligent perception and analysis: The metadata module identifies it as "Sentinel-2 MSI", and the visual analysis module simultaneously detects cloud coverage and longitudinal strip noise in specific bands (such as B01).

[0091] Intelligent Dynamic Decision Making: Faced with complex situations, the decision engine initiates advanced decision-making logic. Its rule base specifies that "classification requires high-quality surface reflectance," and its evaluation model weighs the order and necessity of cloud removal and noise reduction. The final generated workflow is: [Radiometric Calibration] → [Strip Noise Detection and Removal (only for affected bands)] → [Cloud Detection and Labeling] → [Atmospheric Correction] → [Cloud Shadow Restoration (using a deep learning restoration model)] → [Study Area Clipping]. This process is not preset but dynamically generated for the specific combination of "cloud + noise."

[0092] Automated execution and evaluation: The pipeline executes sequentially. The quality module evaluates residual banding after denoising and assesses the coherence of the repaired area with the surrounding texture and spectrum after cloud restoration. If the spectrum of the repaired area is distorted, the feedback signal may trigger a local reprocessing using an alternative restoration algorithm.

[0093] Output: The system outputs a surface reflectance product that is free from stripe noise and cloud cover effects and has been restored to the maximum extent, providing input data for subsequent land cover classification that is far superior to traditional fixed processes (which may only perform atmospheric correction).

[0094] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A remote sensing image intelligent preprocessing method based on a localized multimodal visual model, characterized in that, include: Obtain the metadata of the input remote sensing image and the user's task objective, and parse the image attribute data of the input remote sensing image based on the sensor parameter knowledge base; A lightweight visual analysis model is invoked in the local computing environment to perform a preliminary quality assessment of the input remote sensing image, and preliminary quality judgment data is obtained. Based on the image attribute data, the initial quality assessment data, and the user task objective, multiple remote sensing image preprocessing operations are dynamically retrieved, evaluated, sorted, and assembled from a preset set of remote sensing image preprocessing operations to generate a preprocessing execution chain for the input remote sensing image. The corresponding remote sensing image preprocessing operation is executed according to the preprocessing execution chain, and standardized data flow is performed between each remote sensing image preprocessing operation according to the preset data interface. Intermediate processing results are obtained at key processing nodes of the preprocessing execution chain, and the intermediate processing results are scored for quality. If the quality score does not meet the preset quality conditions, the preprocessing execution chain is adjusted according to the quality score. The adjustment includes at least one of processing parameter adjustment, processing operation replacement, or remedial processing step insertion. The output is a standardized remote sensing image data product processed by the preprocessing execution chain.

2. The intelligent preprocessing method for remote sensing images based on a localized multimodal visual model according to claim 1, characterized in that, The image attribute data includes at least one of the following: data source, sensor type, imaging time, band information, and projection parameters; The initial quality assessment data includes at least one of cloud coverage information, noise information, missing stripe information, image quality overview information, or image texture information.

3. The intelligent preprocessing method for remote sensing images based on a localized multimodal visual model according to claim 1, characterized in that, The step of dynamically retrieving, evaluating, sorting, and assembling multiple remote sensing image preprocessing operations from a preset set of remote sensing image preprocessing operations based on the image attribute data, the preliminary quality assessment data, and the user task objective, to generate a preprocessing execution chain for the input remote sensing image, includes: Based on the image attribute data and the user task objective, determine the basic preprocessing steps from the preset expert rule base; Based on the image attribute data, the initial quality assessment data, and the user task objective, the suitability of candidate preprocessing operations in the remote sensing image preprocessing operation set is evaluated to obtain the necessity weight of each candidate preprocessing operation and the recommended initial processing parameter range. Based on the basic preprocessing steps, the necessity weights, and the recommended initial processing parameter range, the candidate preprocessing operations are sorted and assembled according to their suitability and necessity weights to generate a preprocessing execution chain that includes processing steps, preprocessing operation identifiers, execution order, and initial processing parameters.

4. The intelligent preprocessing method for remote sensing images based on a localized multimodal visual model according to claim 1, characterized in that, The set of remote sensing image preprocessing operations includes multiple operations such as radiometric calibration, atmospheric correction, geometric correction, orthorectification, image fusion, cloud detection and processing, image mosaicking, image cropping, noise suppression, image registration, or image enhancement. The multiple remote sensing image preprocessing operations are determined based on the user's task objective, the image attribute data of the input remote sensing image, and the initial quality assessment data.

5. The intelligent preprocessing method for remote sensing images based on a localized multimodal visual model according to claim 1, characterized in that, The preprocessing execution chain is described as work order data in a preset format that includes multiple processing nodes and node dependencies; The step of performing the corresponding remote sensing image preprocessing operation according to the preprocessing execution chain includes: Construct a directed acyclic graph based on the node dependencies; Perform a topological sort on the directed acyclic graph; Based on the topology sorting results, the remote sensing image preprocessing operations corresponding to multiple processing nodes are scheduled sequentially, and processing nodes that do not have data dependencies on each other are scheduled in parallel. The remote sensing image preprocessing operation is encapsulated in the form of containerized services. The containerized service corresponding to the processing node is called according to the work order data in the preset format, and remote sensing image data, processing parameters and processing status data are transferred between different containerized services through the preset data interface. Record the execution status of each processing node, and control the startup of subsequent processing nodes based on the execution status.

6. The intelligent preprocessing method for remote sensing images based on a localized multimodal visual model according to claim 1, characterized in that, Before calling the lightweight visual analysis model in the local computing environment, a local model deployment process is performed in advance, which includes: The pre-trained model used for preliminary assessment of remote sensing image quality is subjected to at least one of the following processes: model quantization, operator fusion, or computational graph optimization, to obtain a locally deployed model for local inference. The inference backend, computational accuracy, and image segmentation strategy of the locally deployed model are determined based on the CPU, GPU, or memory resources of the local computing device. The input remote sensing images, intermediate processing results, and standardized remote sensing image data products are all processed in a local computing environment and are not sent to the cloud model service.

7. The intelligent preprocessing method for remote sensing images based on a localized multimodal visual model according to claim 1, characterized in that, The key processing nodes include at least one of the following: atmospheric correction node, geometric correction node, cloud detection processing node, or mosaic processing node. The quality scoring of the intermediate processing results includes: The quality assessment sub-model is invoked to quantify and score at least one of the following indicators of the intermediate processing results: radiometric consistency, geometric correction residual, cloud processing integrity, or mosaic seam visibility. The quality assessment sub-model shares the underlying feature extraction network with the lightweight visual analysis model or the artificial intelligence model used to perform at least one remote sensing image preprocessing operation.

8. The intelligent preprocessing method for remote sensing images based on a localized multimodal visual model according to claim 1, characterized in that, The step of adjusting the preprocessing execution chain based on the quality score includes: When the quality score does not meet the preset quality conditions, at least one of the following adjustment methods will be executed: Adjust the processing parameters of the current processing step corresponding to the key processing node, and re-execute the current processing step; Replace the remote sensing image preprocessing operation used in the current processing step with a backup preprocessing operation selected from the preset set of remote sensing image preprocessing operations, and re-execute the current processing step. A remedial processing step is inserted in the preprocessing execution chain at a position following the current processing step.

9. The intelligent preprocessing method for remote sensing images based on a localized multimodal visual model according to claim 1, characterized in that, The output is a standardized remote sensing image data product processed by the preprocessing execution chain, including: The standardized remote sensing image data product is output as a geospatial data file containing georeferenced information. The standardized remote sensing image data product is a geospatial data product that has undergone physical correction and optimization, and does not contain land cover classification results, target detection results, or semantic segmentation results. Generate a processing metadata report associated with the standardized remote sensing image data product. The processing metadata report includes at least one of the following: preprocessing execution chain, processing steps, processing parameters, quality scores of key processing nodes, and execution status. The standardized remote sensing image data product and the processing metadata report are associated and archived.

10. A remote sensing image intelligent preprocessing system based on a localized multimodal visual model, characterized in that, include: The metadata intelligent perception and parsing module is used to acquire the metadata of the input remote sensing image and the user's task objectives, and to parse the image attribute data of the input remote sensing image based on the sensor parameter knowledge base. The local preliminary quality assessment module is used to call a lightweight visual analysis model in the local computing environment to perform a preliminary quality assessment of the input remote sensing image and obtain preliminary quality assessment data. The dynamic decision-making module is used to dynamically retrieve, evaluate, sort, and assemble multiple remote sensing image preprocessing operations from a preset set of remote sensing image preprocessing operations based on the image attribute data, the preliminary quality assessment data, and the user's task objective, and generate a preprocessing execution chain for the input remote sensing image. The pipeline execution module is used to execute the corresponding remote sensing image preprocessing operation according to the preprocessing execution chain, and to perform standardized data flow between each remote sensing image preprocessing operation according to a preset data interface. The quality feedback module is used to obtain intermediate processing results at key processing nodes of the preprocessing execution chain, score the intermediate processing results for quality, and adjust the preprocessing execution chain according to the quality score when the quality score does not meet the preset quality conditions. The adjustment includes at least one of processing parameter adjustment, processing operation replacement, or remedial processing step insertion. The local model deployment module is used to deploy the lightweight visual analysis model in a local computing environment, and to ensure that the input remote sensing images, intermediate processing results and standardized remote sensing image data products are all processed in the local computing environment and are not sent to the cloud model service. The output module is used to output standardized remote sensing image data products after being processed by the preprocessing execution chain.

Images