Method and System for Generating Continuous Cloud-Free Products for Sentinel-2 in a General-Purpose Regional Remote Sensing Production Platform

By preprocessing multi-source remote sensing data, using a six-dimensional spatiotemporal joint index and a conditional diffusion reconstruction network, the problems of low regional scale efficiency, seam artifacts, and insufficient automation in the generation of Sentinel-2 remote sensing products have been solved. This has enabled efficient and seamless remote sensing product generation and standardized output, which is adapted to the business needs of regional platforms.

CN123116145APending Publication Date: 2026-11-17XINJIANG INST OF ECOLOGY & GEOGRAPHY CHINESE ACAD OF SCI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611301288.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-26
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for generating Sentinel-2 remote sensing products suffer from problems such as low efficiency of regional scale inference, poor adaptability of temporal input windows, mosaic seams in parallel reconstruction of sub-blocks, lack of fully automated production lines, and insufficient multi-source fusion and engineering productization, making them unsuitable for the needs of general regional remote sensing platforms.

Method used

A Sentinel-2 continuous cloudless product generation system for a general regional remote sensing production platform was constructed by employing multi-source remote sensing data preprocessing, six-dimensional spatiotemporal joint indexing, conditional diffusion reconstruction network, dynamic time window optimization, distributed scheduling, and seam sensing fusion technology. The system includes modules such as data ingestion, spatiotemporal indexing, dynamic time window selection, conditional diffusion reconstruction, rapid inference, and seam sensing fusion.

Benefits of technology

It achieves efficient inference for regional batch production, adapts the timing input window to complex scenarios, completely eliminates mosaic seam artifacts, builds a fully automated production platform link, significantly improves the accuracy of multi-source data fusion and product standardization, and is compatible with the entire business chain of the regional platform.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116145A_ABST
    Figure CN123116145A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for generating continuous cloudless products using Sentinel-2 for general-purpose regional remote sensing production platforms, belonging to the field of remote sensing AI reconstruction and regional data engineering technology. It includes: standardized preprocessing of multi-source remote sensing data and construction of a six-dimensional spatiotemporal index; solving for the optimal dynamic time window by integrating cloud cover, time interval, seasonal phase, and computing resources; constructing a conditional diffusion network with SAR, time coding, and cloud mask multi-source conditional tensors as input; using an ODE solver for adaptive low-step inference to output tile reconstruction sub-blocks and a confidence field; completing seam sensing fusion based on the confidence field to eliminate tile seams; generating standardized cloudless products with accompanying quality and phylogenetic metadata and releasing them for external service. This invention innovatively couples five mechanisms: dynamic temporal window optimization, conditional decoupling diffusion reconstruction, ODE fast sampling, training + inference two-stage seam constraints, and distributed spatiotemporal scheduling, solving the defects of low regional production efficiency, poor temporal adaptation, artifacts in stitching, and lack of an integrated engineering link.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of remote sensing AI reconstruction and regional data engineering technology, and in particular to a method and system for generating continuous cloudless products for Sentinel-2, a general-purpose regional remote sensing production platform. Background Technology

[0002] The Sentinel-2 optical satellite possesses high-frequency surface observation capabilities, serving as a core data source for dynamic surface monitoring. However, clouds, fog, and terrain shadows can cause large-area observation gaps. With an average global cloud cover of approximately 66%, a single optical time series is insufficient to produce continuous, comprehensive, cloud-free remote sensing products. The Sentinel-1 SAR possesses all-weather cloud-penetrating observation capabilities, supplementing information on sub-cloud features. Existing studies such as RESTORE-DiT and DiffCR have confirmed that conditional diffusion models fusing SAR time series and time-coded data can achieve cloud removal from individual image patches. However, current technologies suffer from five major drawbacks:

[0003] Regional scale inference is inefficient: Traditional diffusion model inference requires hundreds to thousands of iterations of sampling, and regional production needs to handle massive tasks across MGRS tiles, across AOI, and multiple time phases, resulting in extremely high end-to-end processing latency and computing resource consumption; existing low-step sampling algorithms are only for 256×256 small image verification and are not adapted to regional batch production scenarios, and cannot balance sampling throughput and reconstruction accuracy.

[0004] Poor adaptability of time series input windows: Existing solutions use fixed-duration long windows to filter time series observation data. When faced with high cloud cover areas, rapidly changing phenology areas, and irregular observation intervals, they cannot balance the effective number of observations, time series continuity, and computing cost. Although some studies have proposed optimization ideas for time series input of arbitrary length, they have not incorporated cloud cover, phenology, and computing resources into a unified optimization framework, and cannot adapt to the dynamic load of regional platforms.

[0005] Parallel reconstruction of sub-blocks has mosaic seams: regional production uses patch tile block parallel inference, and direct splicing after reconstruction is prone to seam artifacts such as radiometric brightness discontinuity, texture breakage, and spectral discontinuity; existing mosaic tools only smooth seams in the post-processing stage and are not coupled with the generative model training process, so they cannot suppress boundary distortion problems from the source.

[0006] Lack of fully automated regional production lines: Existing cloud-based image removal algorithms are only experimental at the level of single scenes and small image blocks. They have not built an integrated platform link that includes multi-source data aggregation, metadata governance, distributed scheduling, quality control, and product release, and cannot support the production of standardized products in batches across regions.

[0007] There is a gap between multi-source fusion and engineering productization: Existing research only verifies the improvement effect of SAR and time coding on local reconstruction, and has not built a complete engineering solution from multi-source data acquisition, spatiotemporal retrieval, AI reconstruction, seamless fusion to standardized product services. The algorithm model cannot be directly implemented in regional remote sensing production platforms.

[0008] Regional remote sensing data platforms need to possess layered capabilities encompassing "data aggregation—governance—sharing—application" and a data lifecycle management framework. In cross-regional monitoring scenarios such as Central Asia, issues such as data fragmentation and difficulties in cross-agency data collaboration still exist. Existing technologies cannot simultaneously address the five major pain points of efficiency, temporal continuity, stitching consistency, automated production, and multi-source engineered fusion. Therefore, there is an urgent need for an integrated cloud-free product generation solution adapted to general-purpose regional remote sensing platforms. Summary of the Invention

[0009] In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide a method and system for generating Sentinel-2 continuous cloudless products for a general regional remote sensing production platform.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] This invention provides a method for generating continuous cloud-free products using Sentinel-2 for general-purpose regional remote sensing production platforms, comprising the following steps:

[0012] S1. Input Stage: Acquire multi-source remote sensing raw data corresponding to the target area and target temporal range. The multi-source remote sensing raw data should include at least Sentinel-2 optical image sequences and Sentinel-1 SAR image sequences, along with auxiliary quality data. Perform standardized preprocessing on all multi-source remote sensing raw data, including radiometric normalization, geometric registration, coordinate system unification, resolution alignment, outlier removal, and metadata registration, to obtain a standardized dataset. Based on the standardized dataset, construct a six-dimensional spatiotemporal joint index including spatial, temporal, cloud cover, observation coverage, seasonal similarity, and computing resource dimensions.

[0013] S2, Time Series Filtering Processing Stage: Read the cloud cover status C, observation time interval T, surface seasonal status S, and platform computing power resource status R stored in the spatiotemporal joint index, and input the four types of parameters into a unified optimization objective function to solve for the optimal dynamic time window W* that adapts to the current target tile and time phase;

[0014] S3. Conditional Sample Construction Stage: Extract target temporal cloud pollution optical images, all SAR temporal features within the dynamic time window W, temporal cloud masks, temporal coding, tile spatial context quality priors, and splice them to form a multi-source fusion conditional tensor c.

[0015] S4, AI Reconstruction Inference Stage: The conditional tensor c is input into the conditional diffusion reconstruction network trained with joint loss. Based on the ODE solver of the ordinary differential equation, the number of inference steps is adaptively determined to perform low-step diffusion sampling, and the output is a tile-based cloudless reconstruction sub-block and a confidence field generated pixel by pixel. The conditional diffusion reconstruction network adopts a conditional decoupled coding architecture. During the training stage, reconstruction loss, spectral consistency loss, temporal consistency loss, seam loss, and data prediction loss are introduced to form a joint loss constraint.

[0016] S5, Seamless Fusion Stage: Read the cloudless reconstructed sub-blocks and corresponding generated confidence fields of the overlapping areas of adjacent tiles, construct a weighted optimization model to perform seam-aware fusion, complete radiometric consistency correction, edge smoothing, inter-spectral consistency constraint, global seam optimization, eliminate tile splicing artifacts, and output a complete cloudless image at the regional level.

[0017] S6. Product Generation and Output Stage: Based on complete cloudless regional imagery, standardized remote sensing products are generated in batches. These standardized remote sensing products include continuous cloudless reflectance products, RGB true-color products, and remote sensing index products. Observation count layers, quality confidence layers, uncertainty layers, and product genealogy metadata are output simultaneously. Tile tasks for the entire region are distributed in batches through a distributed scheduling mechanism. All standardized remote sensing products are released to the outside world based on RESTful API and OGC standard tile services. The full-process audit metadata of the products is stored synchronously for task backtracking and quality sampling verification.

[0018] Furthermore, the optimization function for solving the optimal dynamic time window W* is:

[0019]

[0020] The objective function can be defined as follows:

[0021]

[0022] The meanings of each item are as follows:

[0023] Ω: Set of all candidate time windows

[0024] : Average residual cloud cover or effective cloud pollution rate within the time window;

[0025] : Weighted time deviation of candidate observations from the target at time phase t0;

[0026] Seasonal / phenological differences can be represented by annual day sequence (DOY), land surface state code, or historical statistical similarity.

[0027] Inference cost under resource state R (GPU, CPU, storage, I / O, task queue) constraints;

[0028] Effective observation availability;

[0029] , , , , These are the weighting coefficients for each item.

[0030] Furthermore, the auxiliary quality data includes at least one or more of the following: cloud mask, scene classification layer, solar zenith angle information, satellite orbit information, terrain-corrected DEM data, and historical surface quality identifiers.

[0031] Furthermore, the forward diffusion process of the conditional diffusion reconstruction network follows a Gaussian distribution:

[0032]

[0033] The original closed-form sampling is expressed as:

[0034]

[0035] The reverse reconstruction process is constrained by the multi-source condition tensor c:

[0036]

[0037] In the formula, For the spread noise scheduling coefficient, , The mean and covariance matrix of the network predictions.

[0038] Furthermore, the conditional decoupling coding architecture splits the conditional tensor into independent observation condition branches and temporal condition branches. After feature encoding is completed by dedicated attention modules, multi-scale temporal feature fusion is performed to avoid confusion of features with different conditional information.

[0039] Furthermore, the joint loss function is expressed as follows:

[0040]

[0041] in: To mitigate the regionally weighted reconstruction losses, a weight coefficient greater than 1 is set for cloud-polluted areas to enhance the restoration of missing ground feature details; To mitigate spectral consistency loss, spectral fidelity is constrained by combining spectral angle distance (SAM) and band-specific L1 deviation. To mitigate the loss of temporal consistency, the stability characteristics of vegetation indices are used to constrain the continuity of temporal phenology without abrupt changes. To account for seam loss, constrain the brightness and gradient texture consistency of the overlapping area of ​​adjacent tiles; To compensate for data prediction loss, the network directly predicts clean, cloudless imagery, replacing the traditional noise prediction paradigm. , , , , These are the weighting coefficients for each loss term.

[0042] Furthermore, the seam-aware fusion is divided into two stages: pre-training constraint and post-inference optimization. In the training stage, the network output tile boundary consistency is pre-constrained based on the seam loss. In the inference stage, global seam optimization is completed based on the generated confidence field, thus suppressing splicing artifacts from both the training and post-processing layers.

[0043] Furthermore, the adaptive inference step calculation model for the ODE solver is as follows:

[0044]

[0045] In the formula, This refers to the number of reasoning steps per tile; The complexity of tile clouds; As an indicator of spatial heterogeneity of the Earth's surface; Cluster computing power congestion level; , The minimum and maximum number of reasoning steps are preset; , , This is the adjustment coefficient; The function is used to restrict the calculation result to [ , Within the range.

[0046] Furthermore, the distributed scheduling mechanism decomposes the global production task into a three-element unit of "spatial tile - dynamic time window - computing node"; the scheduling optimization objective is to minimize the maximum completion time of the cluster and the cross-node data transmission overhead, and the constraints include the upper limit of single node video memory, memory, IO load, and priority deployment of adjacent tile tasks on the same node to reduce the amount of boundary data communication.

[0047] Furthermore, standardized remote sensing products generate complete product line metadata, which records the original data source, dynamic time window parameters, inference steps, fusion weights, cloud mask input, and computing power scheduling logs. When the overall quality confidence layer is lower than the preset threshold, a fallback strategy is automatically triggered to re-execute the windowing screening time series, increase the inference steps, or switch to the medoid time series statistical synthesis fallback scheme.

[0048] This invention also provides a Sentinel-2 continuous cloudless product generation system for a general-purpose regional remote sensing production platform, used to execute the generation method described in any of the above-mentioned methods. The system includes seven functional modules connected in series:

[0049] The data ingestion module is used to acquire Sentinel-2 optical sequences, Sentinel-1 SAR sequences, and auxiliary quality data, perform standardized preprocessing and metadata registration, and output a standardized dataset.

[0050] The spatiotemporal index module is used to build a six-dimensional spatiotemporal joint index based on standardized datasets, and supports joint retrieval based on multiple conditions such as region, time phase, cloud volume, and computing resources.

[0051] The dynamic time window selection module is used to read the four types of parameters C, T, S, and R of the spatiotemporal index and solve for the optimal dynamic time window W* through an optimization function;

[0052] The conditional diffusion reconstruction module is used to construct a multi-source fusion conditional tensor c, and the diffusion network based on the conditional decoupling coding architecture outputs tile-based cloudless reconstruction sub-blocks and generates a confidence field.

[0053] The fast inference module is used to adaptively allocate inference steps based on the ODE solver, and complete low-step diffusion sampling to accelerate inference.

[0054] The seam perception fusion module is used to build a weight optimization model based on the confidence field of the tile overlap area, to complete seamless mosaicking across the entire domain and eliminate splicing seam artifacts.

[0055] The product generation and service output module is used to generate multiple types of standardized remote sensing products in batches, along with supporting quality layers and spectral metadata. It connects to a distributed scheduling cluster, publishes products to the outside world through multiple standard service interfaces, and supports task backtracking and quality sampling verification.

[0056] Compared with the prior art, the technical solutions disclosed in the embodiments of the present invention have the following beneficial effects:

[0057] Significantly improve the inference efficiency of regional batch production

[0058] Leveraging the adaptive low-step sampling mechanism of the S4 ODE solver, the traditional diffusion model's hundreds of iterations are compressed to single digits to twenty steps. Combined with the distributed temporal and spatial desiccation model, the global tile task is split into parallel computations, solving the shortcomings of high computing power consumption and high processing latency at the regional scale in the background technology. The adaptive step strategy distinguishes between simple and complex surface tiles, taking into account both reconstruction accuracy and overall throughput efficiency, and realizing automated batch production over a large area.

[0059] The timing input window adapts to various complex observation scenarios.

[0060] Based on the S2 and CTSR four-parameter joint dynamic time window optimization mechanism, it replaces the traditional fixed-duration time series window and simultaneously balances the four major constraints of cloud cover, phenology, observation interval, and cluster computing power. It solves the problem of insufficient utilization of time series information under high cloud areas, irregular observations, and fluctuations in computing power load. The time series reconstruction results show continuous phenology without seasonal breaks.

[0061] Completely eliminate regional tile parallel reconstruction mosaic seam artifacts

[0062] By employing a dual mechanism of pre-constraining seam loss during the S4 training phase and confidence-driven seam perception fusion during the S5 inference phase, the consistency of tile boundary pixels is constrained in advance during the training phase, and the radiation, spectrum, and texture of the overlapping area are adaptively smoothed based on the generated confidence during the post-processing phase. This addresses the shortcomings of existing technologies that only perform simple smoothing in post-processing and suffer from severe boundary distortion, and outputs a standardized remote sensing product with continuous space and no seams visible to the naked eye.

[0063] Establish an integrated, fully automated regional remote sensing production platform link

[0064] It fully covers the entire process of multi-source data acquisition, spatiotemporal indexing, time-series filtering, AI reconstruction, seamless fusion, and standardized product release, and is equipped with metadata genealogy governance, automated quality control, and task rollback fallback strategies; it makes up for the shortcomings of existing algorithms that can only be tested on small image patches and cannot be engineered and implemented on regional platforms, and can directly connect to the regional remote sensing platform's "collection-governance-sharing-application" layered business framework.

[0065] Deep fusion of multi-source heterogeneous data significantly improves the fidelity of on-premises reconstruction.

[0066] S3 employs a multi-condition decoupled coding input diffusion network based on SAR time series, time coding, cloud masking, and tile context, and is equipped with multi-dimensional joint loss to constrain the accuracy of spectral, temporal, and detail reconstruction. It fully leverages the advantages of SAR cloud penetration observation to effectively model long-term land cover change patterns, and solves the technical defects of texture, spectral distortion, and temporal jumps in the reconstruction of cloud cover from single optical images. The reconstructed land cover details and vegetation phenology are more consistent with the real surface conditions.

[0067] The product is standardized, reusable, and adaptable to the entire business chain of regional platforms.

[0068] In addition to basic cloudless imagery, it simultaneously outputs quality confidence layer, observation counting layer, and full-process spectral metadata, along with OGC and RESTful standard service interfaces; it supports task backtracking, quality sampling, and thematic secondary processing, and can directly support upper-level businesses such as ecological monitoring, agricultural yield estimation, and cross-border regional decision support, thereby enhancing the platform's data reuse value. Attached Figure Description

[0069] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0070] Figure 1 This is a system architecture diagram of the "Sentinel-2 Continuous Cloudless Product Generation" regional remote sensing production platform provided in an embodiment of the present invention.

[0071] Figure 2 This is a data ingestion module architecture diagram provided in an embodiment of the present invention;

[0072] Figure 3 This is a diagram illustrating the architecture of the spatiotemporal indexing module provided in an embodiment of the present invention.

[0073] Figure 4 This is a diagram illustrating the architecture of the dynamic time window selection module provided in an embodiment of the present invention.

[0074] Figure 5 This is a diagram illustrating the architecture of a conditional diffusion reconstruction module provided in an embodiment of the present invention.

[0075] Figure 6 This is a diagram illustrating the architecture of a fast inference module provided in an embodiment of the present invention.

[0076] Figure 7 This is a data mosaic module architecture diagram provided in an embodiment of the present invention;

[0077] Figure 8 This is a product generation module architecture diagram provided in an embodiment of the present invention;

[0078] Figure 9 The algorithm flowchart provided for the embodiments of the present invention. Detailed Implementation

[0079] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0080] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0081] like Figures 1-9As shown, this embodiment of the invention provides a method for generating Sentinel-2 continuous cloudless products for a general-purpose regional remote sensing production platform, which is divided into four main stages:

[0082] Phase 1: Input and Preparation (corresponding to S1)

[0083] Target definition: Enter the target area AOI, target time range, and required product type;

[0084] Multi-source retrieval and ingestion: Retrieve auxiliary data such as Sentinel-2, Sentinel-1, DEM, and cloud masks from remote sensing data archives;

[0085] Construction of a six-dimensional spatiotemporal index: After preprocessing, a joint index of space-time-cloud cover-seasonal phase-observation-resource is constructed.

[0086] Phase Two: Task Planning (corresponding to S2 and S3)

[0087] 4. Dynamic time window solution: Calculate the optimal time window W* based on index parameters C / T / S / R;

[0088] 5. Multiple rounds of data supplementation retrieval: Supplement corresponding temporal optical and SAR observation data a second time according to the window range;

[0089] 6. Conditional Tensor Sample Construction: Encode SAR, time, cloud mask, and tile context, and stitch together the model input conditional tensor.

[0090] Phase 3: Execution and Processing (corresponding to S4 and S5)

[0091] 7. Distributed task scheduling and ODE fast diffusion inference: Split tasks by tile-time window and distribute them to computing power nodes. Adaptively allocate steps to complete reconstruction and output tile sub-blocks and confidence fields.

[0092] 8. Seam-aware fusion: Overlapping area weight optimization, radial correction, edge smoothing, and global seam optimization;

[0093] 9. Regional Product Assembly: Integrate global tiles to generate basic product and quality layers with reflectivity, RGB, and remote sensing index.

[0094] Phase Four: Quality Control and Delivery (corresponding to S6)

[0095] 10. Automated quality control: Read the quality confidence layer to evaluate the accuracy of the entire product. If the accuracy is not met, initiate the window expansion, step increment, statistical synthesis fallback process.

[0096] 11. Standardized Product Release and Retrospective: Output COG / GeoTIFF standard images, share them externally through API and tile services, and store complete metadata for task review and quality sampling.

[0097] The overall technical solution proposed in the embodiments of the present invention can be summarized as follows:

[0098] Data ingestion → Spatiotemporal indexing → Dynamic time window selection → Conditional diffusion reconstruction → ODE / solver rapid inference → Seam-aware fusion → Product generation and service release.

[0099] The data layer references the multi-source access, standardized aggregation, metadata governance, and API service framework of the regional platform; the model layer absorbs RESTORE-DiT's time-series conditional utilization of SAR+date information, as well as DiffCR's diffusion modeling ideas for conditional decoupling, data prediction, and low-step sampling; and the engineering layer draws on the production experience of regional mosaic tools for cross-AOI, cross-tile, and streaming sub-block aggregation, and extends it into a unified platform solution of "generative reconstruction + seamless mosaicking".

[0100] This invention provides a method and system for generating continuous cloudless products using Sentinel-2 for general-purpose regional remote sensing production platforms. The specific purpose of this invention is to address the shortcomings in efficiency, continuity, scalability, and automation in stitching regional-scale cloudless products. By integrating various heterogeneous remote sensing data, such as optical imagery, SAR imagery, and DEM elevation data, the raw data is converted into a standardized representation within the system. Simultaneously, metadata is automatically extracted and processed to establish complete data lineage and quality labels, providing a standardized, high-quality data foundation for subsequent indexing, scheduling, and reconstruction tasks.

[0101] Based on spatiotemporal database technology, a hybrid index structure combining R-tree variants and time partitioning is adopted to support multi-level and multi-resolution spatiotemporal queries, enabling efficient retrieval and on-demand scheduling of massive remote sensing data, and allowing reconstruction tasks to quickly locate available observation data within the target spatiotemporal range.

[0102] For the target region and time range, an optimal time-series data observation window is adaptively generated based on constraints such as data availability, cloud cover, and seasonal consistency. This window is not a fixed-length historical segment, but rather an optimal subsequence dynamically selected by the algorithm, striking a balance between temporal coverage continuity and data quality reliability. This provides sufficient temporal information and reasonable redundancy as data input for subsequent diffusion reconstruction. Using the dynamically selected time-series observation sequence as a condition, the diffusion model is driven to perform generative reconstruction of the missing region. This step employs the following design:

[0103] Temporal condition coding: Drawing on the RESTORE-DiT architecture's use of SAR observations and date information as temporal conditions, time coding and SAR observation quantization are embedded into the temporal attention mechanism of the diffusion model. This allows the model to simultaneously perceive the observations themselves and the temporal location information of "when was what observed", thereby incorporating the seasonal patterns and trend priors of ground feature changes into the reconstruction process.

[0104] Conditional decoupling: Drawing on the concept of conditional decoupling from DiffCR, the reconstruction conditions are explicitly split into two orthogonal components: "existing observation conditions" and "time conditions," which are then input into the model's independent attention modules to avoid information confusion between different conditions.

[0105] Data Prediction Paradigm: This paradigm replaces noise prediction in traditional diffusion models, allowing the model to directly predict the complete reconstruction result at the target time during the diffusion process, rather than predicting noise first and then inferring the reconstruction value. This reduces the accumulation of uncertainty during diffusion and improves reconstruction stability under low-step conditions. It also accelerates the sampling process of conditional diffusion reconstruction to a controllable low-step range. This step utilizes an efficient numerical solver for Ordinary Differential Equations (ODEs) to replace the stepwise denoising sampling of traditional diffusion models. It compresses the diffusion sampling process, which originally required tens to hundreds of steps, into one or several ODE solution steps. While ensuring reconstruction quality, it reduces inference latency by more than an order of magnitude, meeting the processing efficiency requirements of engineering production scenarios. It also seamlessly fuses reconstruction results across AOIs or across tiles. This step not only handles spatial geometric alignment and radiation consistency but also introduces the concept of a "generative confidence field"—quantifying the generative confidence of each pixel during conditional diffusion reconstruction into a spatial field. During fusion, it fully preserves the texture features of high-confidence regions and adaptively smooths the seams of low-confidence regions. This step inherits and expands upon the production experience of the regional mosaic tool for cross-AOI, cross-tile, and streaming sub-block aggregation, but replaces the image tiling operation in traditional stitching with semantic-level seam-aware fusion based on generation confidence, thereby eliminating boundary artifacts while preserving reconstruction details; the reconstruction results after seam-aware fusion are packaged into a standard remote sensing product, equipped with complete cloud mask, quality assessment indicators and metadata information.

[0106] The product is released to the public through RESTful API, OGC standard services or dynamic tile services, supporting on-demand subscription, spatial retrieval and time backtracking, forming an integrated publishing capability from data input to terminal services.

[0107] The technical effects of the above technical solution are as follows:

[0108] Efficiency Improvement: By relying on low-step solver sampling, the cost of diffusion inference is significantly reduced, which helps to compress the diffusion process that originally required hundreds to thousands of steps into a dozen or even a few dozen steps, thus making regional batch production feasible.

[0109] Scalable: Through tile-streaming, cross-AOI / cross-tile processing and distributed scheduling, local reconstruction capabilities can be extended to multi-tile, multi-temporal, and cross-regional continuous production scenarios.

[0110] Enhanced temporal continuity: By utilizing SAR time series, date encoding, dynamic time windows, and time consistency loss, the generated results are made more consistent with the logic of continuous observation, reducing the problems of breaks or seasonal misalignment caused by fixed windows.

[0111] Significantly reduced seams: Through seam loss and seamless weight optimization, better consistency in brightness, texture, and spectral transitions is achieved at region boundaries and overlapping areas, improving the map-level usability of the product.

[0112] Enhanced automation and platform compatibility: This solution is naturally suited for deployment on regional platforms with data governance, shared services, and API capabilities, facilitating subsequent expansion to index products, thematic products, and decision-making services.

[0113] The mathematical model and core algorithm of this invention are as follows:

[0114] Dynamic time window selection model

[0115] Define the target time phase as t0, and the candidate time window set as Ω. The dynamic time window can be represented as:

[0116]

[0117] The objective function can be defined as:

[0118]

[0119] The meanings of each item are as follows:

[0120] : Average residual cloud cover or effective cloud pollution rate within the time window;

[0121] : Weighted time deviation of candidate observations from the target at time phase t0;

[0122] Seasonal / phenological differences can be represented by annual day sequence (DOY), land surface state code, or historical statistical similarity.

[0123] Inference cost under resource state R (GPU, CPU, storage, I / O, task queue) constraints;

[0124] Effective observation availability (e.g., effective pixel ratio, number of candidate phases, SAR / optical pairing integrity);

[0125] , , , , These are the weighting coefficients for each item.

[0126] This formula integrates "cloud cover, temporal consistency, seasonal consistency, resource cost, and effective observation count" into a single optimization framework to address the problem that fixed windows cannot adapt to different regions, seasons, cloud conditions, and computing power levels. Its design is based on publicly available research recognizing the importance of irregular time intervals, date information, and arbitrary length sequence inputs, as well as the practical resource constraints of regional streaming processing.

[0127] To reduce engineering complexity, several candidate windows can be generated in the actual implementation:

[0128]

[0129] Then for each Calculate J(W) i Select the optimal window. .in It can dynamically change within a preset range according to the cloud conditions in the region and the platform load, for example, controlled by strategies such as "short window priority, expanding the window when necessary, and limiting the flow when resources are scarce".

[0130] Conditional diffusion reconstruction model

[0131] Make the target cloudless optical image as The diffusion steps are t=1,…,T. The forward diffusion process is defined as:

[0132]

[0133] It can also be written in closed-form sampling of the original sample:

[0134]

[0135] The reverse reconstruction process is defined as:

[0136]

[0137] In the formula, For the spread noise scheduling coefficient, , The mean and covariance matrix of the network predictions.

[0138] Where cc is the set of conditional information, which includes at least:

[0139] Target temporal cloud contamination optical image ;

[0140] In-window SAR time series features

[0141] cloud mask

[0142] Time coding

[0143] • Optional topographic coding, seasonal coding, and historical statistical coding.

[0144] The mathematical forms of the aforementioned forward / backward diffusion are consistent with the standard definitions in remote sensing diffusion cloud removal methods. The innovation of this invention lies in embedding these definitions into the dynamic time window and sub-block stitching scenarios in the regional production platform.

[0145] Conditional coding can be further represented as:

[0146]

[0147] in For the SAR encoder, g represents the spatial tile context, and q represents the quality prior (e.g., effective pixel ratio, boundary reliability, etc.). Temporal coding can use a sine / cosine formula:

[0148]

[0149] This allows for the simultaneous expression of adjacent time differences and the position of seasonal phases within the year.

[0150] Loss function design

[0151] This invention employs combined loss:

[0152]

[0153] in: To mitigate the regionally weighted reconstruction losses, a weight coefficient greater than 1 is set for cloud-polluted areas to enhance the restoration of missing ground feature details; To mitigate spectral consistency loss, spectral fidelity is constrained by combining spectral angle distance (SAM) and band-specific L1 deviation. To mitigate the loss of temporal consistency, the stability characteristics of vegetation indices are used to constrain the continuity of temporal phenology without abrupt changes. To account for seam loss, constrain the brightness and gradient texture consistency of the overlapping area of ​​adjacent tiles; To compensate for data prediction loss, the network directly predicts clean, cloudless imagery, replacing the traditional noise prediction paradigm. , , , , These are the weighting coefficients for each loss term.

[0154]

[0155] in:

[0156] (1) Reconstruction losses

[0157]

[0158] in The target phase cloud mask is used, with ρ>1 to increase the recovery weight of cloud-contaminated areas. This ensures that clean areas are not over-rewritten while enhancing the recovery of cloud details.

[0159] (2) Loss of spectral uniformity

[0160]

[0161] Where SAM is the spectral angular distance and B is the number of bands. This term is used to constrain the angle and band deviation between the reconstructed spectrum and the true cloudless spectrum, avoiding the problem of "visually smooth but spectrally distorted". S2Mosaic's emphasis on "preserving the true observed spectral shape" in the medoid method also illustrates that regional products cannot only focus on visual effects while ignoring spectral consistency.

[0162] (3) Time consistency loss

[0163]

[0164] in It is a time-stable feature extractor that can output vegetation index, water content index, texture statistics or low-dimensional representation; Determined by temporal distance and seasonal similarity. This term is used to constrain the continuity of the target temporal reconstruction results with the neighboring reference temporal phases in terms of phenological and temporal changes, avoiding "jumps" in the temporal sequence after reconstruction.

[0165] (4) Joint loss

[0166] For adjacent sub-blocks Overlapping area mask definition:

[0167]

[0168] The first term constrains the consistency of brightness / reflectivity in the overlapping area, while the second term constrains the smoothness of gradients and texture boundaries. This term is a new training objective added to this invention for region-level patch / tile parallel inference, used to mitigate the risk of subsequent mosaicking seams in advance during the training phase.

[0169] (5) Data prediction items

[0170] A better approach is to "predict clean data rather than predict noise":

[0171]

[0172] Public experiments show that, when there is sufficient conditional information, dataprediction converges more easily than noiseprediction, especially when the number of multi-temporal conditional inputs increases. This provides a direct basis for the multi-source conditional diffusion of the present invention.

[0173] ODE / Solver Accelerated Inference Model

[0174] Rewrite the backdiffusion sampling problem as an ODE solving problem:

[0175]

[0176] A single update can then be uniformly represented as:

[0177]

[0178] in The solver update operator can be a DPM-Solver, a first-order / second-order / third-order dedicated solver, or other high-order ODE samplers. Existing research shows that the bottleneck of traditional diffusion sampling is the excessive number of steps, while the DPM-Solver compresses sampling to 10–20 function evaluations; DiffCR further demonstrates that under specific diffusion design conditions, one-step generation and 3–5-step convergence can be achieved. Based on this, this invention adopts a resource-aware adaptive step-counting strategy: dynamically allocating the sampling step count K according to cloud complexity, tile size, and resource state R.

[0179] Defineable:

[0180]

[0181] in:

[0182] The target is the complexity of the tile cloud;

[0183] It serves as an indicator of surface heterogeneity.

[0184] Resource congestion;

[0185] This refers to the number of reasoning steps per tile;

[0186] Cluster computing power congestion level;

[0187] , The minimum and maximum number of reasoning steps are preset;

[0188] , , This is the adjustment coefficient;

[0189] The function is used to restrict the calculation result to [ , Within the range.

[0190] This allows for fewer steps of reasoning in stable regions and more steps in complex regions, thus balancing speed and quality.

[0191] Distributed scheduling model (spatiotemporal resource decomposition)

[0192] The regional production task is recorded as:

[0193]

[0194] Where G represents the set of tiles in the regional space. The target time phase set. The design resource node set is as follows. The resource vector for each node is:

[0195]

[0196] The scheduling objective is to minimize the maximum completion time and data migration overhead.

[0197]

[0198] The constraints include:

[0199] 1. Each task is executed by only one node;

[0200] 2. Task execution must meet constraints related to video memory, main memory, and I / O;

[0201] 3. Overlapping boundary tasks of adjacent tiles can be prioritized for co-location to reduce boundary communication;

[0202] 4. Quality rollback tasks should prioritize reusing the original cache to avoid repeated reads.

[0203] This model breaks down regional production into a three-element combination of spatial blocks, time windows, and resource nodes, representing a key engineering innovation that truly grounds diffusion and reconstruction on the regional production platform.

[0204] Seamless mosaic optimization model

[0205] Suppose there are n candidate sub-block reconstruction results in the overlapping region. fusion results for:

[0206]

[0207] The optimization objective is defined as:

[0208]

[0209] in:

[0210] The overlapping area;

[0211] Reference statistics for the overlapping area (e.g., median / robust mean);

[0212] To mitigate losses;

[0213] The quality penalty term can consist of factors such as cloud residual probability, boundary uncertainty, and insufficient observations.

[0214] This optimization ensures that weights change continuously in space and prioritizes higher-quality local reconstruction results, thus achieving a seamless transition rather than abrupt cropping. Its engineering motivation aligns with the requirements of region-level tile-streaming, cross-AOI aggregation, and output mesh consistency.

[0215] The algorithm flow mainly includes:

[0216] 1. Target definition: Input target region Ω, target time phase set T, and product type set p.

[0217] 2. Multi-source retrieval and ingestion: Retrieve Sentinel-2, Sentinel-1 and auxiliary quality layers, and complete standardization preprocessing.

[0218] 3. Construct a spatiotemporal index: Establish indexes for spatial tiles, time, cloud cover, effective observations, resource status, and product lineage.

[0219] 4. Dynamic time window solution: Solve for the optimal time window W for each target tile / phase.

[0220] 5. Conditional sample construction: Extracting cloud pollution optical images, SAR features, cloud masks, time coding, and contextual quality priors.

[0221] 6. Task decomposition and scheduling: Distribute tasks to the distributed execution queue according to the "spatial block - time window - resource node" structure.

[0222] 7. Low-step conditional diffusion reasoning: The number of steps K is adaptively determined based on resource status and complexity, and ODE / solver reasoning is performed.

[0223] 8. Overlapping area seam perception fusion: Perform weight optimization and boundary smoothing on the overlapping areas of sub-blocks.

[0224] 9. Regional-level product assembly: Generate cloudless reflectivity, RGB, index products and quality layers.

[0225] 10. Quality Control and Rollback: If the quality layer is below the threshold, trigger window expansion, step increment, supplementary candidates, or switch rollback strategies.

[0226] 11. Publishing and Retrospection: Output GeoTIFF / COG / service interfaces and write genealogy and audit metadata.

[0227] The innovative aspects of this invention are as follows:

[0228] 1. Dynamic Time Window Joint Optimization Mechanism: For the first time, cloud cover C, time interval T, seasonal phase S, and resource status R are unified into a regional production-level time window optimization function W=f(C,T,S,R), which changes the time window selection from a "fixed parameter" to an adaptive process driven by "data and resources".

[0229] 2. Regional platform-based conditional diffusion reconstruction architecture: Sentinel-1SAR, cloud mask, date encoding and temporal context are used as conditional inputs. Regional tile context and quality prior are introduced to enable diffusion reconstruction to directly serve regional product production, rather than just single scene / patch level repair.

[0230] 3. Solver-driven low-step diffusion inference: Couples ODE / solver sampling with regional resource scheduling, and adaptively adjusts the number of sampling steps according to task complexity to achieve a balance between regional throughput and quality.

[0231] 4. Dual mechanism of training-phase seam loss + inference-phase seam optimization: Not only does it perform seamless stitching in post-processing, but it also constrains boundary consistency in advance during the training phase through seam loss, reducing tile gaps from the source.

[0232] 5. Distributed scheduling model for spatiotemporal resource decomposition: The regional production task is divided into a three-element task unit of spatial block, time window and resource node, and a scheduling strategy coupled with the generation model is constructed.

[0233] Product-level output and pedigree governance are integrated: outputting quality confidence layers, observation count layers, and pedigree metadata in addition to images, enabling them to be directly integrated into the regional platform's sharing, distribution, and reuse system.

[0234] Example 1: Production of Sentinel-2 reflectivity products in a continuous cloudless region across Central Asia

[0235] 1) Applicable Scenarios

[0236] This approach aims to generate continuous cloud-free reflectance products on a decadal or monthly scale for Central Asia, a region encompassing multiple MGRS tiles, administrative units, or transnational borders. This implementation is particularly suitable for ecologically fragile, fragmented, and inter-agency collaborative regions like Central Asia.

[0237] 2) Parameter Examples

[0238] Spatial processing units: 1024×1024 to 4096×4096 pixel tiles;

[0239] Overlap width: 32 to 256 pixels;

[0240] Output projection: The target area's UTM or a specified equal-area projection;

[0241] Candidate time window range: 5–90 days before and after the target time phase;

[0242] Number of sampling steps: 3–20 steps adaptive selection;

[0243] SAR conditions: At least one of VV and / or VH polarization must be introduced;

[0244] Cloud mask: SCL or deep learning cloud mask can be used;

[0245] Output: COG / GeoTIFF + quality layer + observation count layer.

[0246] 3) Process

[0247] 1. Retrieve Sentinel-2 and Sentinel-1 data around the target time phase in the target area from the STAC / archive system;

[0248] 2. Register optical and SAR data to a unified projection and a unified pixel grid;

[0249] 3. Calculate the cloud cover, effective pixel rate, number of candidate observations, seasonal bias, and resource cost for each tile;

[0250] 4. Solve for W* in time phase for each tile / target;

[0251] 5. Construct a conditional tensor c and input it into a conditional diffusion network;

[0252] 6. Using a low-step solver, the reconstruction results and uncertainties are obtained through reasoning;

[0253] 7. Seamlessly merge and tile adjacent sub-blocks;

[0254] 8. Generate the product and add it to the database, then output the API and download service.

[0255] 4) Implementation Points

[0256] When the target area has extremely high cloud cover and insufficient effective optical observations, the time window should be extended first.

[0257] When resources are congested, prioritize using lower steps for areas with low heterogeneity;

[0258] When the joint quality score is below the threshold, perform incremental re-pushing or window expansion re-pushing on the boundary reconstruction block.

[0259] Example 2: Ten-day Continuous Products in Central Asian Agricultural Monitoring Scenarios

[0260] This embodiment focuses on continuous monitoring in agricultural areas, with particular attention to crop seasonal phases and index continuity. The system is an improvement upon Embodiment 1. Agricultural index features, such as NDVI, EVI, and water content indicators, are extracted to emphasize the smoothness and reasonable variation of crop time series, thus mitigating temporal consistency loss. Dynamic time windows prioritize candidate observations with minimal seasonal differences from the target time phase, allowing the resource module to increase inference steps when necessary to achieve higher temporal consistency. Public research has demonstrated the significant value of date information and SAR time series in restoring crop growth status; this embodiment elevates this to part of the regional decadal-scale product logic.

[0261] Example 2: Ten-day Continuous Products in Central Asian Agricultural Monitoring Scenarios

[0262] Central Asian agricultural region.

[0263] Experimental steps:

[0264] 1. Retrieve Sentinel-2 and Sentinel-1 data around the target time phase in the target area from the STAC / archive system;

[0265] 2. Agricultural index feature extraction, such as NDVI, EVI, and water content indicators, emphasizes the smoothness and reasonable variation of crop time series to mitigate temporal consistency loss.

[0266] 3. Register optical and SAR data to a unified projection and a unified pixel grid;

[0267] 4. Calculate the cloud cover, effective pixel rate, number of candidate observations, seasonal bias, and resource cost for each tile;

[0268] 5. Solve for W* for each tile / target in time phase;

[0269] 6. Construct a conditional tensor c and input it into a conditional diffusion network;

[0270] 7. Using a low-step solver, the reconstruction results and uncertainties are obtained through reasoning.

[0271] 8. Seamlessly merge and mosaic adjacent sub-blocks;

[0272] 9. Generate the product and add it to the database, then output the API and download service.

[0273] This invention also provides a Sentinel-2 continuous cloudless product generation system for a general-purpose regional remote sensing production platform, used to execute the generation method described in any of the above embodiments. The system includes seven functional modules connected in series:

[0274] The data ingestion module is used to acquire Sentinel-2 optical sequences, Sentinel-1 SAR sequences, and auxiliary quality data, perform standardized preprocessing and metadata registration, and output a standardized dataset.

[0275] The spatiotemporal index module is used to build a six-dimensional spatiotemporal joint index based on standardized datasets, and supports joint retrieval based on multiple conditions such as region, time phase, cloud volume, and computing resources.

[0276] The dynamic time window selection module is used to read the four types of parameters C, T, S, and R of the spatiotemporal index and solve for the optimal dynamic time window W* through an optimization function;

[0277] The conditional diffusion reconstruction module is used to construct a multi-source fusion conditional tensor c, and the diffusion network based on the conditional decoupling coding architecture outputs tile-based cloudless reconstruction sub-blocks and generates a confidence field.

[0278] The fast inference module is used to adaptively allocate inference steps based on the ODE solver, and complete low-step diffusion sampling to accelerate inference.

[0279] The seam perception fusion module is used to build a weight optimization model based on the confidence field of the tile overlap area, to complete seamless mosaicking across the entire domain and eliminate splicing seam artifacts.

[0280] The product generation and service output module is used to generate multiple types of standardized remote sensing products in batches, along with supporting quality layers and spectral metadata. It connects to a distributed scheduling cluster, publishes products to the outside world through multiple standard service interfaces, and supports task backtracking and quality sampling verification.

[0281] The specific module functions are as follows:

[0282] Data Ingestion Module

[0283] The data collected for the target area at the target time phase and its adjacent time phases includes at least the following:

[0284] Sentinel-2 optical imaging sequence;

[0285] Sentinel-1 SAR image sequence;

[0286] Cloud mask / scene classification layer, solar zenith angle, orbital information, terrain assistance layer, quality labeling layer;

[0287] Optional auxiliary data include meteorological data, land surface classification data, historical data, and manually verified data.

[0288] This module performs radiometric normalization, geometric registration, coordinate unification, resolution alignment, outlier removal, and metadata registration on multi-source data to form data objects that can be scheduled by the platform.

[0289] Spatiotemporal Index Module

[0290] A spatial tile index, a time index, a quality index, and a resource index are constructed for the data to support joint queries by region, time phase, cloud cover, observation coverage, seasonal similarity, and available resources.

[0291] Spatiotemporal Index Module Architecture Diagram

[0292] This module corresponds to the "data catalog, metadata management, quality control, hierarchical management, data distribution and API interface" capabilities in the internal platform.

[0293] Dynamic Time Window Selection Module

[0294] The time window W is adaptively determined based on cloud cover status C, time interval T, seasonal phase status S, and resource status R, replacing the fixed monthly / ten-day window scheme.

[0295] This module addresses the balance between "expanding the window when observations are insufficient, shrinking the window when seasonal deviations are too large, and reducing the number of candidates when resources are scarce," and is one of the core innovations of this invention.

[0296] Conditional diffusion reconstruction module

[0297] The module takes the target temporal cloud-contaminated optical image as the main input and Sentinel-1SAR features, temporal coding, cloud mask, and temporal context as conditional inputs. It outputs the cloudless reconstruction result under the target temporal cloud through a conditional diffusion network. This module adopts a conditional decoupled coding and temporal conditional fusion structure to reduce training and inference complexity and enhance temporal consistency.

[0298] Fast Inference Module

[0299] The diffusion inversion process is rewritten as a low-step sampling process compatible with ODE / solver. By using DPM-Solver or other high-order solvers, the number of inference steps is compressed from the traditional hundreds / thousands of steps to single digits to tens of steps, thereby meeting the throughput requirements of regional batch production.

[0300] Seam sensing fusion module

[0301] In overlapping regions of adjacent sub-blocks, radiometric consistency correction, edge smoothing, inter-spectral consistency constraints, and seam optimization are performed to achieve seamless mosaicking at the region level. This module not only fuses the reconstruction results in the post-processing stage but also pre-constrains the boundary consistency of the model output through seam loss during training.

[0302] Product Generation Module

[0303] The output should include at least:

[0304] Continuously cloudless Sentinel-2 reflectivity products;

[0305] RGB / True Color Products;

[0306] Index products (such as NDVI, NDWI, EVI, etc.);

[0307] Observation counting layer, quality confidence layer, uncertainty layer, product lineage / source tracing metadata;

[0308] Service interfaces (download, API, task replay, quality sampling).

[0309] This module is integrated with the "Shared Service Layer, Application Service Layer, Data Distribution, Special Applications, and Decision Support" in the internal platform.

[0310] The innovations are as follows:

[0311] 1. Dynamic Time Window Joint Optimization Mechanism: For the first time, cloud cover C, time interval T, seasonal phase S, and resource status R are unified into a regional production-level time window optimization function W=f(C,T,S,R), which changes the time window selection from a "fixed parameter" to an adaptive process driven by "data and resources".

[0312] 2. Regional platform-based conditional diffusion reconstruction architecture: Sentinel-1SAR, cloud mask, date encoding and temporal context are used as conditional inputs. Regional tile context and quality prior are introduced to enable diffusion reconstruction to directly serve regional product production, rather than just single scene / patch level repair.

[0313] 3. Solver-driven low-step diffusion inference: Couples ODE / solver sampling with regional resource scheduling, and adaptively adjusts the number of sampling steps according to task complexity to achieve a balance between regional throughput and quality.

[0314] 4. Dual mechanism of training-phase seam loss + inference-phase seam optimization: Not only does it perform seamless stitching in post-processing, but it also constrains boundary consistency in advance during the training phase through seam loss, reducing tile gaps from the source.

[0315] 5. Distributed scheduling model for spatiotemporal resource decomposition: The regional production task is divided into a three-element task unit of spatial block, time window and resource node, and a scheduling strategy coupled with the generation model is constructed.

[0316] 6. Product-level output and pedigree governance integration: Output quality confidence layer, observation count layer and pedigree metadata in addition to imagery, so that they can be directly entered into the regional platform's sharing, distribution and reuse system.

[0317] Alternative implementation methods and scalability

[0318] Sensors are replaceable:

[0319] Sentinel-2 can be replaced with Landsat-8 / 9, GF series, or commercial high-resolution optical sources; Sentinel-1 can be replaced with SAR sources such as GF-3, RADARSAT, or ALOS. As long as the principle of "optical main mode + cloud penetration auxiliary mode + temporal index" is met, it can be included in the framework of this invention.

[0320] The model is replaceable:

[0321] Conditional diffusion networks can be replaced by generative models such as DDPM, DDIM, LatentDiffusion, Score-basedModel, RectifiedFlow, and ConsistencyModel, as long as the core mechanism of "multi-source conditional constraints + low-step high-throughput + regional seam consistency" is retained.

[0322] Deployment mode can be replaced:

[0323] It can be deployed in public clouds, private clouds, hybrid clouds, HPC clusters, edge computing nodes close to the data, or "cloud-edge collaboration" architectures. The basic capabilities of cloud computing platforms, distributed storage, network communication, and security protection embodied in the internal platform materials are all suitable as a hosting environment.

[0324] Product types can be expanded:

[0325] In addition to continuous cloudless reflectance, it can also be extended to true color products, index products, monthly composite products, anomaly detection base maps, disaster pre- and post-disaster comparison base maps, training sample base maps, and spatiotemporal thematic services.

[0326] The fallback strategy is scalable:

[0327] When the confidence level of diffusion reconstruction is low, it can automatically switch to traditional synthesis strategies such as medoid, median, and first-valid, forming a dual-track production mode of "generative priority and statistical fallback". S2Mosaic has demonstrated the trade-offs between speed, smoothness, and spectral consistency of different synthesis methods, and this invention incorporates them into the platform fallback strategy.

[0328] The basic principles of the present invention have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in the present invention are merely examples and not limitations, and should not be considered as essential features of each embodiment of the present invention. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the present invention to the necessity of employing the aforementioned specific details.

[0329] The block diagrams of devices, apparatuses, devices, and systems involved in this invention are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0330] It should also be noted that in the apparatus, device, and method of the present invention, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of the present invention.

[0331] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the invention. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of the invention. Therefore, the invention is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0332] It should be understood that the qualifying terms "first", "second", "third", "fourth", "fifth" and "sixth" used in the description of the embodiments of the present invention are only used to more clearly illustrate the technical solutions and are not intended to limit the scope of protection of the present invention.

[0333] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of the invention to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A method for generating continuous cloudless products using Sentinel-2 for a general-purpose regional remote sensing production platform, characterized in that, Includes the following steps: S1. Input stage: Acquire multi-source remote sensing raw data corresponding to the target area and target temporal range. The multi-source remote sensing raw data shall include at least Sentinel-2 optical image sequence, Sentinel-1 SAR image sequence, and supporting auxiliary quality data. Standardized preprocessing was performed on all multi-source remote sensing raw data, including radiometric normalization, geometric registration, coordinate system unification, resolution alignment, outlier removal, and metadata registration, to obtain a standardized dataset. A six-dimensional spatiotemporal joint index is constructed based on standardized datasets, including spatial, temporal, cloud cover, observation coverage, seasonal similarity, and computing resources dimensions. S2, Time Series Filtering Processing Stage: Read the cloud cover status C, observation time interval T, surface seasonal status S, and platform computing power resource status R stored in the spatiotemporal joint index, and input the four types of parameters into a unified optimization objective function to solve for the optimal dynamic time window W* that adapts to the current target tile and time phase; S3. Conditional Sample Construction Stage: Extract target temporal cloud pollution optical images, all SAR temporal features within the dynamic time window W, temporal cloud masks, temporal coding, tile spatial context quality priors, and splice them to form a multi-source fusion conditional tensor c. S4, AI Reconstruction Inference Stage: The conditional tensor c is input into the conditional diffusion reconstruction network trained with joint loss. Based on the ODE solver of the ordinary differential equation, the number of inference steps is adaptively determined to perform low-step diffusion sampling, and the output is a tile-based cloudless reconstruction sub-block and a confidence field generated pixel by pixel. The conditional diffusion reconstruction network adopts a conditional decoupled coding architecture. During the training stage, reconstruction loss, spectral consistency loss, temporal consistency loss, seam loss, and data prediction loss are introduced to form a joint loss constraint. S5, Seamless Fusion Stage: Read the cloudless reconstructed sub-blocks and corresponding generated confidence fields of the overlapping areas of adjacent tiles, construct a weighted optimization model to perform seam-aware fusion, complete radiometric consistency correction, edge smoothing, inter-spectral consistency constraint, global seam optimization, eliminate tile splicing artifacts, and output a complete cloudless image at the regional level. S6. Product Generation and Output Stage: Based on complete cloudless regional imagery, standardized remote sensing products are generated in batches. These standardized remote sensing products include continuous cloudless reflectance products, RGB true-color products, and remote sensing index products. Observation count layers, quality confidence layers, uncertainty layers, and product genealogy metadata are output simultaneously. Tile tasks for the entire region are distributed in batches through a distributed scheduling mechanism. All standardized remote sensing products are released to the outside world based on RESTful API and OGC standard tile services. The full-process audit metadata of the products is stored synchronously for task backtracking and quality sampling verification.

2. The generation method according to claim 1, characterized in that, The optimization function for solving the optimal dynamic time window W* is: The objective function can be defined as follows: The meanings of each item are as follows: Ω represents the set of all candidate time windows. : Average residual cloud cover or effective cloud pollution rate within the time window; : Weighted time deviation of candidate observations from the target at time phase t0; Seasonal / phenological differences can be represented by annual day sequence, land surface state code, or historical statistical similarity. The reasoning cost under resource state R constraints; Effective observation availability; , , , , These are the weighting coefficients for each item.

3. The generation method according to claim 1, characterized in that, The auxiliary quality data includes at least one or more of the following: cloud mask, scene classification layer, solar zenith angle information, satellite orbit information, terrain-corrected DEM data, and historical surface quality identifiers.

4. The generation method according to claim 1, characterized in that, The forward diffusion process of the conditional diffusion reconstruction network follows a Gaussian distribution: The original closed-form sampling is expressed as: The reverse reconstruction process is constrained by the multi-source condition tensor c: In the formula, For the spread noise scheduling coefficient, , The mean and covariance matrix of the network predictions.

5. The generation method according to claim 1, characterized in that, The joint loss function is expressed as follows: in: To mitigate the regionally weighted reconstruction losses, a weight coefficient greater than 1 is set for cloud-polluted areas to enhance the restoration of missing ground feature details; To mitigate spectral consistency loss, spectral fidelity is constrained by combining spectral angle distance (SAM) and band-specific L1 deviation. To mitigate the loss of temporal consistency, the stability characteristics of vegetation indices are used to constrain the continuity of temporal phenology without abrupt changes. For seam loss, the brightness and gradient texture consistency of the overlapping area of ​​adjacent tiles are constrained; Ldata is the data prediction loss, and the network directly predicts clean, cloudless images to replace the traditional noise prediction paradigm. , , , , These are the weighting coefficients for each loss term.

6. The generation method according to claim 1, characterized in that, The seam-aware fusion is divided into two stages: pre-training constraints and post-inference optimization. In the training stage, the network output tile boundary consistency is pre-constrained based on the seam loss. In the inference stage, global seam optimization is completed based on the generated confidence field, thus suppressing splicing artifacts from both the training and post-processing layers.

7. The generation method according to claim 1, characterized in that, The adaptive inference step calculation model for the ODE solver is as follows: In the formula, This refers to the number of reasoning steps per tile; The complexity of tile clouds; As an indicator of spatial heterogeneity of the Earth's surface; Cluster computing power congestion level; , The minimum and maximum number of reasoning steps are preset; , , This is the adjustment coefficient; The function is used to restrict the calculation result to [ , Within the range.

8. The generation method according to claim 1, characterized in that, The distributed scheduling mechanism decomposes the global production task into a three-element unit of "spatial tile - dynamic time window - computing node"; the scheduling optimization objective is to minimize the maximum completion time of the cluster and the cross-node data transmission overhead, and the constraints include the upper limit of single node video memory, memory, IO load, and priority deployment of adjacent tile tasks on the same node to reduce the amount of boundary data communication.

9. The generation method according to claim 1, characterized in that, Standardized remote sensing products are equipped with complete product genealogy metadata, which records the original data source, dynamic time window parameters, inference steps, fusion weights, cloud mask input, and computing power scheduling logs. When the overall quality confidence layer is lower than the preset threshold, a fallback strategy is automatically triggered to re-execute the windowing screening time series, increase the inference steps, or switch to the medoid time series statistical synthesis fallback scheme.

10. A Sentinel-2 continuous cloudless product generation system for a general-purpose regional remote sensing production platform, characterized in that: For executing the generation method according to any one of claims 1-9, the system comprises seven functional modules connected in series: The data ingestion module is used to acquire Sentinel-2 optical sequences, Sentinel-1 SAR sequences, and auxiliary quality data, perform standardized preprocessing and metadata registration, and output a standardized dataset. The spatiotemporal index module is used to build a six-dimensional spatiotemporal joint index based on standardized datasets, and supports joint retrieval based on multiple conditions such as region, time phase, cloud volume, and computing resources. The dynamic time window selection module is used to read the four types of parameters C, T, S, and R of the spatiotemporal index and solve for the optimal dynamic time window W* through an optimization function; The conditional diffusion reconstruction module is used to construct a multi-source fusion conditional tensor c, and the diffusion network based on the conditional decoupling coding architecture outputs tile-based cloudless reconstruction sub-blocks and generates a confidence field. The fast inference module is used to adaptively allocate inference steps based on the ODE solver, and complete low-step diffusion sampling to accelerate inference. The seam perception fusion module is used to build a weight optimization model based on the confidence field of the tile overlap area, to complete seamless mosaicking across the entire domain and eliminate splicing seam artifacts. The product generation and service output module is used to generate multiple types of standardized remote sensing products in batches, along with supporting quality layers and spectral metadata. It connects to a distributed scheduling cluster, publishes products to the outside world through multiple standard service interfaces, and supports task backtracking and quality sampling verification.