A geological large model construction method and system based on video processing

Abstract

The application discloses a geological large model construction method and system based on video processing, comprising: acquiring a geological data set; using a U-Net of lithology filling area semantic segmentation to perform visual analysis on geological maps to obtain a lithology filling area mask; using a text detection network and a text recognition network cascade model to perform OCR and field structured extraction on scanned documents and map them to a geological data dictionary; using a space-time Transformer to perform space-time feature coding on a deformation sequence of a remote sensing and InSAR time sequence video to generate a space-time embedding representation; constructing a cross-modal sample unit and aligning it to a unified geological object primary key index to form a traceable training sample; and based on the sample, training a geological basic large model containing a video encoder to generate a target task result through cross-modal fusion and a task output head. The application improves multi-source alignment efficiency and reasoning reviewability, and greatly improves the accuracy of geological data processing.

Description

Technical Field

[0001] This invention relates to the field of video processing technology, and more specifically to a method and system for constructing a large geological model based on video processing. Background Technology

[0002] With the rapid growth in demand for urban underground space development, major engineering projects, and geological disaster prevention, urban-scale geological informatization is gradually shifting from the traditional "map-report-manual interpretation" model to a "multi-source sensing-spatiotemporal analysis-intelligent decision-making" model. Urban geological data sources are extremely complex, including structured data such as geological map raster images, scanned images from exploration documents, and borehole logging, as well as high-frequency spatiotemporal data such as surface change videos acquired through remote sensing and time-series deformation videos generated by InSAR monitoring. These data cover a spatial scale ranging from point boreholes to regional remote sensing, and a temporal scale ranging from one-time explorations to continuous monitoring. The data formats span images, videos, text tables, and structured attributes, posing significant challenges to data fusion, unified modeling, and interpretable applications.

[0003] In existing technologies, the digitization of geological maps and exploration data often employs a combination of scanning and data entry, manual vectorization, and manual input. While this approach can achieve some level of data archiving, the raster maps contain a mix of elements such as lithological fill areas, symbol annotations, and border legends. Manual drawing is inefficient and its consistency is subject to subjective influences. Extracting tables and fields from scanned documents often relies on general OCR or manual verification, making it difficult to consistently output field-level structured records that can be directly mapped to the geological data dictionary. This, in turn, affects the standardized data entry and subsequent analysis of key elements such as borehole data, stratigraphy, and lithological descriptions. For remote sensing and InSAR monitoring data, existing processes typically handle these separately within dedicated pipelines for remote sensing change detection and InSAR deformation trend analysis. This lack of a unified indexing and semantic alignment mechanism between geological maps, exploration texts, and borehole structured data makes it difficult to form a cross-modal, traceable sample system, thus limiting the learning and reasoning of the relationships between surface changes, subsurface geological structures, and engineering disturbances.

[0004] In recent years, deep learning has made progress in image segmentation, text recognition, and temporal modeling. Examples include using networks like U-Net for feature or region segmentation, cascading text detection and recognition networks for document OCR, and using Transformers for spatiotemporal sequence encoding. However, these algorithms are mostly deployed in a single-task, single-modal manner: image segmentation results often remain at the mask level, lacking primary key-level associations with borehole layer sequences, deformation sequences, and document fields; OCR outputs lack structured field quality assessment and traceability; and spatiotemporal encoding lacks cross-attention alignment and fusion with geological map semantics and engineering survey elements. Therefore, existing technologies struggle to construct large-scale geological models for urban scales and cannot simultaneously meet the comprehensive requirements of "automated structuring of multi-source heterogeneous data," "cross-modal spatiotemporal alignment and unified indexing," "traceable training sample construction," and "interpretable, auditable, and verifiable results for target tasks." In particular, they struggle to provide evidence chains such as source video clips, timestamps, image regions, and document paragraphs in the output results to support business decisions and accountability.

[0005] For the reasons mentioned above, there is an urgent need for a method and system for constructing large geological models based on video processing. This system should be able to uniformly process, align, and fuse geological map raster, exploration scan documents, remote sensing and InSAR time-series videos, and borehole structured data to form traceable training samples and train a basic geological model. This would output reliable target task results and their evidence chains, providing more automated, consistent, and interpretable technical support for urban geological surveys, monitoring and early warning, and underground space planning. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention discloses a method and system for constructing a large geological model based on video processing. The method includes: acquiring an urban geological data set containing raster images of geological maps, scanned images of exploration documents, remote sensing videos, InSAR time-series videos, and borehole structured data; using U-Net for semantic segmentation of lithological filling areas to perform visual analysis of geological elements on the geological maps, obtaining a lithological filling area mask; employing a cascaded model of text detection and text recognition networks to perform OCR and field structure extraction on the scanned documents and map them to a geological data dictionary; using a spatiotemporal Transformer to encode the deformation sequences of remote sensing and InSAR time-series videos to generate spatiotemporal embedding representations; constructing cross-modal sample units and aligning them to a unified geological object primary key index to form traceable training samples; training a large geological model containing a video encoder based on the samples, generating the target task result through cross-modal fusion and a task output header, and simultaneously outputting an evidence chain containing the primary key index, source video segment identifier and timestamp, source image region location, and source document paragraph location. This invention improves multi-source alignment efficiency and reasoning verifiability.

[0007] This invention provides a method for constructing a large geological model based on video processing, comprising the following steps:

[0008] S1: Obtain a set of urban geological data, which includes: geological map raster images, scanned images of exploration documents, remote sensing videos, InSAR time-series videos, and borehole structured data;

[0009] S2: Perform geological feature visual analysis on the geological map raster image using U-Net with semantic segmentation of lithological filling area. The output geological feature analysis results include: lithological filling area mask;

[0010] S3: The cascaded model of text detection network and text recognition network is used to perform image text recognition and field structure extraction on the scanned image of the exploration document, outputting field-level structured records, and mapping the field-level structured records to a preset geological data dictionary;

[0011] S4: The remote sensing video and InSAR time series video are processed by a spatiotemporal Transformer to extract spatiotemporal features from the deformation sequence and encode them to generate a spatiotemporal embedding representation;

[0012] S5: Construct cross-modal sample units, aligning geological element analysis results, field-level structured records, borehole structured data, and spatiotemporal embedding representations to a unified geological object primary key index, forming traceable training samples;

[0013] S6: Train a large geological foundation model based on traceable training samples. The large geological foundation model includes a multimodal encoder, a cross-modal fusion backbone network, and a task output head, wherein the multimodal encoder includes a video encoder.

[0014] S7: The geological foundation large model outputs the target task results and the evidence chain corresponding to the target task results. The evidence chain includes the primary key index, the source video segment identifier and timestamp, the source image region location, and the source document paragraph location.

[0015] Preferably, the U-Net for semantic segmentation of the lithological filling area in step S2 is an improved U-Net, which includes a multi-scale context encoding branch and a boundary refinement branch; wherein, the multi-scale context encoding branch extracts lithological texture features of different receptive fields through dilated convolution, and the boundary refinement branch enhances the boundary of the lithological filling area through an edge-guided attention module to reduce omissions and misclassifications at geological interfaces.

[0016] Preferably, the improved U-Net sets a cross-layer feature alignment gating module at the skip connection. The cross-layer feature alignment gating module generates gating weights based on the correlation between the features of the encoding end and the decoding end, and adaptively fuses the low-level texture features and high-level semantic features, thereby improving the segmentation robustness of the lithological filling area under the interference of complex symbols and annotations.

[0017] Preferably, step S2 further includes performing layout structure analysis on the geological map raster image, identifying the lithological filling area, legend area, annotation area and border area, and masking the legend area and annotation area to reduce the interference of non-lithological areas on the generation of the lithological filling area mask.

[0018] Preferably, the field structure extraction in step S3 includes table structure recovery. The table structure recovery obtains the table topology by detecting cell boundaries and inferring row and column relationships, and matches the cell content with the field template to generate borehole number, coordinates, elevation, layer depth, and lithological description.

[0019] Preferably, the spatiotemporal Transformer in step S4 adopts a dual-stream coding structure, including a remote sensing video change coding stream and an InSAR temporal deformation coding stream; wherein, the change coding stream encodes the temporal change features of the remote sensing video, the deformation coding stream encodes the deformation sequence of the InSAR temporal video, and the change features and deformation features are aligned and fused through cross-attention modules to generate the spatiotemporal embedding representation.

[0020] Preferably, the alignment to the unified geological object primary key index in step S5 includes: establishing a correlation between the spatial range of the video clip and the geological map raster image, and between the borehole location and the InSAR location based on a joint matching strategy of spatial neighborhood and time window; wherein the spatial neighborhood is determined by the buffer radius, and the time window is determined by the sampling interval or the duration of the event.

[0021] Preferably, the multimodal encoder in step S6 includes a structured data encoder, which encodes the layered sequence and lithological properties in the borehole structured data into a sequence embedding, and aligns it with the spatiotemporal embedding representation output by the video encoder through a cross-modal fusion backbone network to learn the joint representation between the lithological filling mask-borehole layer-deformation sequence.

[0022] Preferably, the evidence chain output in step S7 further includes uncertainty information and counterfactual explanation information. The uncertainty information includes confidence scores or confidence intervals, and the counterfactual explanation information includes key frame indexes, key image region masks and their corresponding document field references that contribute the most to the target task results, thereby supporting the auditing and review of the output results.

[0023] This application also provides a geological large model construction system based on video processing, including:

[0024] The data access module acquires a set of urban geological data, which includes: geological map raster images, scanned images of exploration documents, remote sensing videos, InSAR time-series videos, and borehole structured data.

[0025] The geological element visual analysis module performs geological element visual analysis on the geological map raster image using U-Net with semantic segmentation of lithological filling area. The output geological element analysis results include: lithological filling area mask.

[0026] The field-level structured record extraction module performs image text recognition and field structure extraction on the scanned image of the exploration document using a cascaded model of a text detection network and a text recognition network, outputs field-level structured records, and maps the field-level structured records to a preset geological data dictionary;

[0027] The spatiotemporal embedding representation generation module uses a spatiotemporal Transformer to extract spatiotemporal features from the deformation sequence of the remote sensing video and InSAR time-series video and encodes them to generate spatiotemporal embedding representations.

[0028] The cross-modal fusion module constructs cross-modal sample units, aligning geological element analysis results, field-level structured records, borehole structured data, and spatiotemporal embedding representations to a unified geological object primary key index, forming traceable training samples;

[0029] The training module trains a large geological foundation model based on traceable training samples. The large geological foundation model includes a multimodal encoder, a cross-modal fusion backbone network, and a task output head, wherein the multimodal encoder includes a video encoder.

[0030] The output module outputs the target task results from the geological foundation large model, and outputs the evidence chain corresponding to the target task results. The evidence chain includes the primary key index, the source video segment identifier and timestamp, the source image region location, and the source document paragraph location.

[0031] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0032] (1) This invention incorporates the most difficult-to-process unstructured carriers of urban geological data (geological map raster images, survey document scan images, remote sensing videos, and InSAR time-series videos) into a unified processing chain, realizing an automated closed loop from acquisition and parsing to data entry. By using the lithological filling area semantic segmentation U-Net to output a lithological mask, the efficiency bottleneck and subjective differences caused by traditional manual vectorization and manual interpretation are significantly reduced. By cascading text detection networks and text recognition networks, document field-level structured extraction is achieved and mapped to a preset geological data dictionary, so that core elements such as borehole number, coordinates, elevation, layer depth, and lithological description can be deposited into computable data in a consistent format. Compared with the existing "scanning and archiving + manual entry" method, this invention significantly improves the data entry speed, field consistency, and reusability, providing a stable and high-quality data foundation for subsequent model training and business analysis.

[0033] (2) This invention uses a unified geological object primary key index as a hub to align lithological masks, document field records, borehole structured data, and spatiotemporal embedding representations to construct traceable cross-modal training samples, thus solving the pain point of "each modality is processed separately and it is difficult to link and reason" in the prior art. By encoding the deformation sequence of remote sensing video and InSAR time series video through spatiotemporal Transformer, stable trends and abnormal patterns can be extracted from high-frequency spatiotemporal data, and jointly modeled with geological map semantics and borehole layer sequence in the cross-modal fusion backbone network. After introducing the video encoder, the model can simultaneously learn the correlation between surface changes, deformation evolution and underground geological structure, thereby improving the generalization ability and robustness of target tasks (such as risk identification, change interpretation, model update, etc.) and reducing inference distortion caused by data sparsity, noise or modality missing.

[0034] (3) This invention generates an evidence chain corresponding to the output of the target task results, realizing a closed-loop traceability of "conclusion-evidence-source". The evidence chain clearly includes the primary key index of the geological object, the identifier and timestamp of the source video segment, the location of the source image area and the location of the source document paragraph, enabling business personnel to quickly trace back to the specific video segment, map area and document field, complete cross-validation and accountability tracking, and significantly improve the credibility and auditability of engineering decisions. Furthermore, through the consistent indexing of the evidence chain and multimodal samples, the system can support the location of abnormal outputs: distinguishing between abnormal video deformation, map semantic parsing deviation or document field extraction error, thereby forming an operable correction path. Compared with the traditional "black box model that only gives conclusions" approach, this invention is more suitable for high-requirement scenarios such as urban geological monitoring and early warning, planning approval and risk assessment. Attached Figure Description

[0035] Figure 1 This is a flowchart of a method for constructing a large geological model based on video processing according to the present invention;

[0036] Figure 2 This is a structural diagram of a geological large model construction system based on video processing according to the present invention. Detailed Implementation

[0037] Those skilled in the art will understand that, in order to make the above-mentioned objects, features, and beneficial effects of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This application illustrates a method for constructing a large geological model based on video processing, including the following steps:

[0038] S1: Obtain a set of urban geological data, which includes: geological map raster images, scanned images of exploration documents, remote sensing videos, InSAR time-series videos, and borehole structured data;

[0039] In one embodiment, the process of acquiring and organizing urban geological data sets is as follows:

[0040] 1) Implementation scenario and scope: The target area is the central urban area of ​​a certain city and its surrounding new districts. The engineering scope polygon boundary (AOI) is established according to administrative divisions and geological and geomorphological units. The system acquires data related to geological surveys, engineering construction, and geological disaster monitoring within the AOI, forming an urban geological data set.

[0041] 2) Data Sources and Data Types: The urban geological data set in this embodiment includes at least the following five types of data, each type of data is entered into the database according to the format "Data Source—Original Format—Key Metadata":

[0042] A. Sources of geological map raster images: Scanned historical geological maps, geological survey results maps, engineering geological zoning maps, geological hazard susceptibility zoning maps, etc. Original formats: TIFF / GeoTIFF, PNG, JPG, embedded images in PDFs, etc.; if it is a PDF, extract the page images first or convert it to TIFF. Key metadata: Map sheet number / scale, mapping time, coordinate system (e.g., CGCS2000 / local coordinates), resolution, map border range (boundary coordinates or corner points), legend information, scan DPI. Storage organization: Hierarchical directory structured as "AOI—Map Sheet—Scale—Year" with database records, and file hashes are recorded to prevent duplication.

[0043] B. Sources of scanned images for exploration documents: Paper / electronic scans of exploration reports, borehole columnar sections, geotechnical test reports, construction records, as-built documentation, etc. Original formats: PDF, TIFF, JPG / PNG; can contain multiple pages, multiple orientations, and different resolutions. Key metadata: Project name, client, exploration unit, report number, preparation date, page number, scanning angle, OCR language type. Storage organization: Split into page-level objects according to "Project—Report—Page Number," while preserving the original PDF and page image mapping relationship (page_id ↔ pdf_offset).

[0044] C. Remote Sensing Video Sources: UAV inspection videos, oblique photography flight videos, ground patrol / vehicle driving record videos, or video streams synthesized from multi-temporal remote sensing image sequences. Original Formats: MP4 / H.264, MOV, TS; if synthesized from multi-temporal images, encapsulate in "frame=temporal image" format. Key Metadata: Acquisition timestamp, frame rate, resolution, sensor information, trajectory information (GNSS / IMU or flight path file), camera intrinsic parameters (focal length, distortion parameters). Spatiotemporal Reference: UTC time or local time (including time zone identifier) ​​is uniformly adopted; the spatial reference is consistent with AOI (e.g., CGCS2000). Supplementary Processing: If trajectory information is missing, it is marked as "weak localization" for subsequent temporal change feature encoding or pose estimation via visual odometry.

[0045] D. InSAR Time-Series Video Source: Time-series deformation results (points or rasters) obtained from PS-InSAR or SBAS-InSAR processing, output by an external monitoring platform / thematic processing chain. Original Format: Deformation time-series point table (CSV / GeoPackage), deformation raster sequence (GeoTIFF sequence), and time-series video (MP4 / TS) generated from the raster sequence encapsulation. Key Metadata: Observation date sequence, reference point information, unwrapping parameters, deformation unit (mm), point density, spatial resolution, quality indicators (coherence coefficient / residual). Encapsulation Rules (Generating "Time-Series Video"): Each deformation raster is mapped to a frame, with frame number corresponding to date; simultaneously, an index table of "frame_index ↔ date" is saved for evidence chain tracing.

[0046] E. Structured borehole data, sourced from: exploration databases, geological survey databases, historical Excel / Access / SQL databases, or structured tables extracted from S3 OCR fields. Data table structure: Borehole table: borehole_id, engineering coordinates X / Y, borehole elevation, final borehole depth, opening date, closing date; Layer table: layer_id, borehole_id, top / bottom depth, lithology code, lithology description, age / stratigraphic unit; Test table: sample_id, borehole_id, depth, index values ​​(water cut, density, liquid limit, etc.). Key metadata: coordinate system, depth datum (borehole / elevation datum), field version number, data source identifier.

[0047] 3) Unified Identification and Indexing (Prepared for S5 "Primary Key Index Alignment"): The system generates a traceable primary key for each type of data and establishes cross-modal associations: geo_map_id: geological map raster primary key (map sheet number + year + hash); doc_page_id: document page primary key (report number + page number + hash); rs_video_id / rs_clip_id / rs_frame_id: remote sensing video / snippet / frame primary key; insar_video_id / insar_frame_id: InSAR time-series video and frame primary keys, associated with date; borehole_id / layer_id: borehole and layer primary keys (source system ID or re-encoding ID). Three types of indexes are established: Spatial index: R-tree / GeoHash, enabling fast spatial retrieval of map extent, video trajectory, InSAR points / raster, and borehole locations; Temporal index: enabling time window retrieval of video timestamps and InSAR date series; Project / Project index: archiving documents, boreholes, and videos by project dimension.

[0048] 4) Quality Inspection and Rollback Strategy ("Availability Guarantee" from S1 output): Inbound validation is performed on objects entering the urban geological data set: Format validation: whether it can be decoded, page number integrity, and frame index continuity; Spatiotemporal validation: whether the coordinate system is convertible, and whether timestamps are missing / abnormal; Consistency validation: monotonicity of borehole layer top and bottom depths, and completeness rate of required fields; Rollback strategy: Videos without coordinates are marked as "weakly localized," still usable for temporal coding, but not for spatial alignment; Missing dates in InSAR are marked as missing measurements in the frame-date index table; Tilt / blurred document pages are recorded with quality marks for stronger preprocessing or manual review in the S3 stage.

[0049] 5) The final output of S1, the "Urban Geological Data Set", should include at least the following: a set of geological map raster images and their geo_map_id and spatial extent; a set of survey document scan images and their doc_page_id and field templates; a set of remote sensing videos and their rs_video_id, timestamps (optional), and trajectories; a set of InSAR time-series videos and their insar_video_id, frame-date index, and deformation units; and a borehole structured data table and its borehole_id / layer_id.

[0050] S2: Perform geological element visual analysis on the geological map raster image using U-Net with lithological filling area semantic segmentation. The output geological element analysis results include: lithological filling area mask;

[0051] In some embodiments, the generation of a lithology-filled area mask for a geological map raster image includes the following steps:

[0052] 1) Input data and target: Input a geological map raster image (GeoTIFF / PNG / JPG / TIFF exported from a PDF page). The image typically includes lithological fill areas, geological boundaries, annotation text, symbol points, legend areas, borders, and coordinate grids. Output a lithological fill area mask (a binary or multi-class mask of the same size as the input, where "1" represents lithological fill area pixels and "0" represents non-lithological fill area pixels; if lithological categories are required, it can be expanded to multiple types of masks, but this embodiment mainly uses the "lithological fill area" binary mask.

[0053] 2) Image preprocessing and slicing, color / brightness normalization: White balance correction and histogram truncation are performed on the scanned images (removing yellowing and shadows from the background). Denoising and sharpening: Bilateral filtering or non-local mean denoising is used for low-quality scanned images; light sharpening can be applied to fine line areas to preserve boundary texture. Initial layout screening: "Legend area, annotation area, and border area" are quickly located through border detection / connected component analysis, and suppression masks are generated; pixels in these areas are zeroed or weighted down before network input (consistent with dependent claim 4) to reduce interference from non-lithological areas. Slicing strategy: The entire image is sliced ​​into fixed windows (e.g., 512×512 or 1024×1024), and adjacent slices are overlapped (e.g., 25%) to facilitate the processing of ultra-large image sizes and reduce edge truncation errors.

[0054] 3) Training data annotation (supervision signal) Annotation object: Lithological filling area (area filling area), excluding text annotations, symbol points, borders, and legends. Annotation method: Manual vector drawing or semi-automatic (color clustering / threshold pre-segmentation followed by manual correction) generation of pixel-level labels; labels are cropped synchronously with the slices.

[0055] 4) Model Inference and Mask Stitching: Input the slices into the "Lithomorphic Filling Area Semantic Segmentation U-Net" to obtain the lithological filling area probability map for each slice. Threshold the probability map to obtain a binary mask; then, morphological closing operations / small connected component removal are used to eliminate noise and holes. Overlapping regions are stitched together using weighted fusion (e.g., high weight for the center, low weight for the edges) to output the lithological filling area mask for the entire image. The specific structure of the Lithomorphic Filling Area Semantic Segmentation U-Net (an improved U-Net) improves segmentation accuracy, especially at the boundaries, in scenarios with "dense geological map symbols, complex textures, thin boundary lines, and multi-scale filling blocks," and enhances robustness against annotation / legend interference.

[0056] The overall framework adopts a U-Net encoder-decoder structure, comprising: Encoder: progressively downsamples to extract multi-scale texture and semantic features; Decoder: progressively upsamples to restore spatial resolution; Skip Connection: fuses features from the encoder and decoder at the same scale; Seg Head: outputs a 1×1 convolutional probability map of the lithological filling area. Each stage of the encoder contains "convolutional block + downsampling": Convolutional block: Conv(3×3) → BN → ReLU → Conv(3×3) → BN → ReLU; Downsampling: MaxPool(2×2) or stride=2 convolution; The number of channels can be increased progressively in {64, 128, 256, 512} (a lightweight backbone can also be used instead, but the U-Net style is maintained here). A multi-scale context encoding branch is introduced at the bottleneck layer (deepest feature layer) to cover lithological filling textures and large-area morphologies at different scales. Multiple sets of dilated convolutions are used to extract different receptive field features in parallel (e.g., dilation rate 2 / 4 / 8 three-branch). The parallel branch outputs are concatenated along the channel dimension and then compressed and fused by a 1×1 convolution to obtain the "context-enhanced bottleneck feature". This approach maintains sensitivity to both large lithological filling areas (low-frequency textures) and small fragment filling areas (high-frequency textures), improving cross-scale adaptability. A boundary refinement branch + edge-guided attention is added at the high-resolution stage of the decoder (e.g., the last two stages). The boundary branch input is high-resolution features from the decoder; the boundary prediction head outputs a "boundary probability map" after several 3×3 convolutions; the edge-guided attention module uses the boundary probability map, normalized by Sigmoid, as attention weights to perform pixel-wise weighting on the features of the main segmentation branch. During training, boundary supervision can be added to the boundary branches (edges are automatically extracted as weak supervision by the labeled mask). During inference, the boundary branches participate in feature enhancement but ultimately still output the lithology mask. This significantly reduces "adhesion / fracture / rough edges" at geological interfaces and the edges of filled areas.

[0057] The cross-layer feature alignment gating module adds gating alignment at each skip connection: it takes encoder and decoder features of the same scale; calculates the correlation between the two (which can be reduced by 1×1 convolution and then dot product / channel correlation) to obtain a gating map; uses the gating map to filter encoder features; and then concatenates / adds them with decoder features to enter subsequent convolutional blocks. Function: When the image contains a large number of symbols, annotations, or noisy textures, the gating module can automatically reduce the influence of irrelevant low-level textures in skip connections, improving the segmentation robustness under "complex symbol and annotation interference". 6) Output and thresholding strategy, main output: lithological filling area probability map (H×W×1). Thresholding: a threshold (e.g., 0.5) is selected based on the validation set to obtain a binary mask; an adaptive threshold can be used near the boundary (the threshold is lowered where the boundary confidence is high) to reduce missed segments.

[0058] The improved U-net structure has the following beneficial effects:

[0059] 1) Improved segmentation accuracy, especially significantly reduced omissions / misclassifications at the boundaries. Multi-scale contextual branches introduce large receptive field information through dilated convolution, preventing large lithological filling areas from being misclassified as background due to weak texture. Boundary refinement branches and edge-guided attention enhance the discriminative features near the boundaries, avoiding adhesion or gaps between lithological and non-lithological areas at the boundary, thereby significantly improving boundary IoU and overall mask connectivity.

[0060] 2) More robust to interference from symbols, annotations, and legends. In geological maps, symbol points, text annotations, and legend textures often overlap with lithological filling textures. Traditional U-Net easily misclassifies symbols or text into lithological areas. The cross-layer feature alignment gating module can adaptively filter skip connection information based on the correlation between features at the encoding and decoding ends, suppressing texture channels inconsistent with lithological semantics, thus maintaining stable segmentation even in areas with dense symbols and annotation coverage. 3) Provides higher-quality visual semantic anchors for subsequent cross-modal alignment and large model training. The output lithological filling area mask has more accurate spatial range and boundary morphology, which can be used as "visual key anchors" for subsequent alignment with borehole layers, document fields, and InSAR / remote sensing spatiotemporal embeddings in S5, reducing mismatches and noise propagation during cross-modal sample construction, and improving the training convergence speed and task inference consistency of the geological foundation large model.

[0061] S3: The cascaded model of text detection network and text recognition network is used to perform image text recognition and field structure extraction on the scanned image of the exploration document, outputting field-level structured records, and mapping the field-level structured records to a preset geological data dictionary;

[0062] In some embodiments, the OCR and field structure extraction of scanned images of exploration documents, taking geological exploration data of a certain city as an example, are performed. The data is mainly in PDF format and includes: cover information page, table of contents page, borehole record table page, layered columnar section page, geotechnical test table page, groundwater level record page, etc. The system renders the PDF page by page into scanned images before proceeding to this step. Input is a scanned image of a single page of exploration document (page ID has been generated), with a resolution between 200–400 dpi equivalent. It may contain tilt, shadows, stamp obstruction, broken table lines, and a mixture of handwritten and machine-printed text. Output is a field-level structured record, preferably a "field record table" or "JSON record". Each record includes at least: field code, field value, field type, field confidence level, page location, text box position, and row / column / paragraph position; and completes the mapping and standardization of the preset geological data dictionary.

[0063] Page preprocessing (improving detection and recognition stability) performs the following steps on the input page in sequence: Tilt correction: Identifies the main direction of the page (such as the direction of table lines or text lines) and rotates the page to horizontal; Shadow removal and background normalization: Reduces uneven grayscale caused by yellowing, shadows, and wrinkles; Contrast enhancement and noise reduction: Enhances the visibility of small text and broken table lines and suppresses scanning noise; Area cropping and header / footer removal (optional): Locates and removes headers, footers, and binding hole areas to avoid interference.

[0064] Page layout (determining subsequent extraction strategy): The pre-processed page is analyzed and divided into the following areas: Table area: an area with a clear row and column structure (drilling record table, test table, etc.); Main text area: paragraph-style text area (project overview, survey conclusions, etc.); Chart area: bar charts / section diagrams, etc.; Legend / Signature area: seals, signatures, legend descriptions, etc. Specifically, the table area uses "cell-level extraction," the main text area uses "row-level extraction + field rule matching," and the chart area may only save the OCR text as supplementary evidence (structured format not mandatory).

[0065] Table structure restoration (the key to field structuring) is performed on each table area as follows: Table line detection and repair: Detect horizontal and vertical lines, connect broken line segments, and fill gaps; Grid construction: Construct a table grid based on the intersection points of line segments to obtain the bounding box of each cell; Header recognition: Recognize the header cells of the first row or key columns to obtain a set of field names (such as "hole number, borehole elevation, bottom layer depth, lithological description", etc.); Cell OCR: Perform cascaded OCR (detect → recognize) on each data cell to obtain the cell text and confidence score; Field generation: Combine "header text + cell text + row and column position" into candidate field records. For example, if the "hole number" column in the 3rd row is recognized as "ZK-013", generate a candidate field: field name = hole number, field value = ZK-013, source = the coordinates of this cell.

[0066] Text area field extraction (applicable to project name, date, unit, etc.): Perform the following on the text area: Use cascading OCR to obtain text organized by line; use field templates for matching, such as patterns like "Project Name:", "Report Number:", "Survey Date:", etc.; form candidate records for the matched field names and values, and record their text box positions and line numbers on the page.

[0067] Field mapping to a preset geological data dictionary: A preset geological data dictionary is established, which includes at least: field code, Chinese field name, alias set, data type, unit, legal range, enumeration table, regular expression style (optional), and priority rules. The mapping process is executed in the following order: Field name unification: "borehole number / drill hole number / drill hole ID" are uniformly mapped to the same field code (e.g., borehole_id); Field value typeification: Numeric fields are converted to numeric types; Date fields are unified to the same format; Coordinate fields are unified in units and decimal places; Unit and enumeration unification: e.g., "elevation (m)" is unified to meters; Lithological descriptions are unified to lithological codes according to the lithological dictionary; Consistency verification: for example, the layer depth must meet the requirement that "the layer top depth is less than the layer bottom depth"; Coordinates must fall within the AOI range; Confidence fusion and labeling: The field-level confidence is formed by combining OCR confidence, table structure consistency, and field template matching score; Fields below the threshold are marked as "to be reviewed".

[0068] The final output (field-level structured record style) includes at least the following for each field-level structured record: field code, field value, data type, unit (if any), field-level confidence level; source page ID, source area type (table / body), and source text box position; if from a table: row number, column number, and header text; if from body: paragraph / line number and trigger template.

[0069] The overall cascaded model consists of a "text detection network" and a "text recognition network," which are connected in series: first, the location of the text is determined (detection), and then the text is identified (recognition). The detection network outputs the boundary positions of the text regions, while the recognition network outputs the character sequence and confidence score for each text region. Finally, the recognition results are backfilled into the original page coordinate system, providing reliable positioning for field structuring.

[0070] The text detection network consists of three parts: a feature extraction backbone, whose input is a page image or a local area image of the page. The backbone network extracts features step-by-step through multiple convolutional layers, preserving details of small-sized text while also extracting structural semantic information such as paragraphs / tables, outputting multi-scale feature layers (high-resolution layers for small text, low-resolution layers for large text). A multi-scale feature fusion module fuses feature layers of different scales, enabling the network to simultaneously detect small, dense text and large headings / wide character blocks. A unified fused feature map is obtained after fusion. The detection output head predicts two types of results on the fused feature map: a text region confidence map, representing the probability that each location belongs to a text region; and a text boundary geometric description, representing the boundary shape of the text region, preferably a rotated rectangle or a four-point polygon, thus adapting to tilted scanning, rotated tables, and irregular text blocks. The detection network output is a set of text regions, each region containing its boundary location (rotated box / polygon) and confidence score. Post-detection processing involves removing low-confidence regions and deduplicating overlapping regions to obtain the final set of text boxes.

[0071] The text recognition network consists of four parts: A text region normalization module receives each text region output by the detection network, crops it, and performs perspective and orientation correction to "straighten" tilted text and standardize its height to a fixed value, avoiding recognition instability due to scale differences. A sequence feature extraction module performs convolutional feature extraction on the normalized text image, outputting feature maps that characterize strokes and glyph structures, and organizes the features into a sequenced feature along the text writing direction (enabling subsequent character-by-character reading). A sequence modeling module performs contextual modeling on the sequenced features, learning the dependencies between characters to enhance robustness against overlapping, missing strokes, and noise occlusion. This module can be implemented using a bidirectional sequence modeling unit or an attention-based sequence modeling unit. A character decoding module converts the modeled sequence features into a character sequence output and provides recognition confidence. The decoding module outputs the final text string, which can include character-level or sequence-level confidence scores. The recognition network outputs: the recognition string and confidence score for each text box. The cascaded output to the structured extraction interface cascaded model finally outputs a "collection of text fragments with coordinates". Each text fragment contains at least: text box position (rotated box / polygon coordinates), recognition string, recognition confidence; page ID and page area (table area / body area); if it is in the table area: it can be associated with cell coordinates, thereby directly generating "field name - field value".

[0072] To ensure mapping to the preset geological data dictionary, the following three-tiered rules are preferred for the "explicit rules": Field alias merging rule: synonymous field names are mapped to the same field code; Type and unit rule: data types and units are bound to field codes and forcibly converted; Consistency check rule: fields such as depth, coordinates, and date meet logical constraints, otherwise confidence is reduced or marked for review.

[0073] S4: The remote sensing video and InSAR time series video are processed by a spatiotemporal Transformer to extract spatiotemporal features from the deformation sequence and encode them to generate a spatiotemporal embedding representation;

[0074] In some embodiments, the spatiotemporal embedding representation generation process of remote sensing video and InSAR time-series video is as follows:

[0075] 1) Input Data and Target Input A: Remote sensing video can be UAV inspection video, aerial oblique image video, vehicle / ground inspection video, or a video stream formed by encapsulating multi-temporal remote sensing images in chronological order. Each frame contains an acquisition timestamp and preferably has coarse positioning information (trajectory, attitude, or coordinates that can be registered to a map). Input B: InSAR temporal video is a video stream formed by encapsulating an InSAR temporal deformation raster sequence. Each frame corresponds to a deformation raster map of an observation date and has a "frame index-date" mapping table, and includes quality information such as coherence coefficient / residual (which can be used as an additional channel or accompanying metadata). Output: Spatiotemporal embedding representation forms one or both of the following outputs: 1) Regional spatiotemporal embedding: Represents the changes and deformation patterns of a geological object / region over a period of time; 2) Location-level spatiotemporal embedding: Represents the deformation evolution characteristics of each grid / patch in the time dimension, used for subsequent alignment to a unified geological object primary key index.

[0076] 2) Frame-level preprocessing and spatiotemporal alignment: Spatial registration and resampling unify remote sensing video frames and InSAR frames to the same coordinate system and raster resolution (e.g., unified grid size and coverage), outputting a pixel-alignable frame sequence; for remote sensing frames that cannot be precisely registered, geo-correction or image registration using feature matching is preferred. Deformation proxy sequence construction (for remote sensing video): Since remote sensing video itself may not directly provide deformation, this embodiment converts it into a "deformation proxy sequence." Specific methods include: performing dense matching / optical flow estimation on adjacent frames to obtain a two-dimensional displacement field; performing change detection on multi-temporal remote sensing frames and converting them into a "displacement / settlement indication intensity map"; using displacement amplitude, direction, and change intensity as the "deformation proxy channel" of the remote sensing frame. Quality masking and noise suppression (for InSAR): A quality mask is generated using coherence coefficients, residuals, or effective observation markers; low-coherence areas are masked or downweighted to prevent noise from entering the spatiotemporal coding. Time axis alignment and window slicing: A time axis is constructed based on the observation date sequence. Long sequences are sliced ​​into fixed window lengths (e.g., by week / month or by frame), with each window serving as a spatiotemporal sample input into the model; missing frames are recorded using "missing markers" so that the model can distinguish between "unobserved" and "no variation".

[0077] 3) Input Representation (Multi-channel Frame Representation): The input frames at each time step are preferably organized into multi-channel tensors, including but not limited to: remote sensing deformation proxy channels (displacement amplitude / intensity of change / direction encoding); InSAR deformation value channels (e.g., subsidence or deformation rate); quality channels (coherence coefficient, residuals, effective observation markers); auxiliary channels (optional, such as slope, aspect, land cover category) to improve robustness to observation biases caused by differences in surface material. 4) Output Embedding Generation Method: After model output, a spatiotemporal embedding representation is formed. Typical output methods include: Global convergence embedding: converging the spatiotemporal features of the entire window to obtain a regional vector (used for target task classification, risk scoring, etc.); Grid / Patch embedding: outputting a temporally aggregated vector for each spatial patch (used for subsequent S5 alignment to the geological object primary key index and construction of cross-modal samples).

[0078] The overall structure of the Spatiotemporal Transformer consists of the following basic layers: Spatiotemporal segmentation and embedding layers divide each frame into multiple spatial patches of a fixed size; multi-channel pixels of each patch generate "patch embeddings" through linear mapping; spatial location encoding (representing the patch's position in the grid) and temporal location encoding (representing the frame's order in the sequence) are added; quality encoding (encoding quality channels or quality markers as "reliability hints") is added. The Spatiotemporal Transformer encoder consists of stacked layers, each including: a spatial self-attention sublayer: attention is applied to all patches within the same temporal frame, learning spatial structure and neighborhood relationships (e.g., deformation banding distribution, boundary transitions); a temporal self-attention sublayer: attention is applied to patch sequences at the same spatial location across time, learning temporal evolution patterns (e.g., slow settling, abrupt acceleration, periodic fluctuations); a feedforward network sublayer: nonlinear transformation is applied to the attention output to improve expressive power; residual connections and normalization ensure stability during deep training. The embedding output head can output a global convergent embedding (for region-level tasks) or the embedding of each patch (for position-level alignment and subsequent fusion). Spatial and temporal models are modeled separately to avoid computational explosion caused by applying attention to "all spatiotemporal tokens" all at once; through quality coding and missing data labeling, the model can distinguish between "real, unchanging" and "observational missing / low coherence".

[0079] In some embodiments, an improved spatiotemporal Transformer structure is adopted. The improved structure of the "dual-stream aligned gated spatiotemporal Transformer" is as follows: dual-stream encoding + cross-attention alignment + time-delay gated fusion. The improved model includes two parallel encoding streams and an alignment and fusion module: the remote sensing change encoding stream takes as input a remote sensing deformation surrogate sequence (channels such as displacement / change intensity / direction). This stream focuses on capturing "visible change cues," such as crack propagation signs, surface morphology changes, and texture changes caused by construction disturbances. The InSAR deformation encoding stream takes as input an InSAR deformation sequence and a quality channel. This stream focuses on capturing "quantified deformation evolution," such as changes in settlement rate, abrupt change points, and trend reversals. The cross-attention alignment module introduces cross-attention at the middle and upper levels of the two streams: using the spatiotemporal features of the InSAR stream as a query, it retrieves the most relevant change evidence from the remote sensing stream; simultaneously, using the features of the remote sensing stream as a query, it retrieves the deformation pattern that best explains the change from the InSAR stream; thus achieving bidirectional alignment of "change-deformation."

[0080] The time-delay adaptive gating fusion module considers that the observation times of remote sensing and InSAR are not completely synchronized, and that there is a lag between visible changes and deformation responses on the surface. To address this, a time-delay gating is implemented: alignment weights are generated based on the timestamp difference between the two types of data, the persistence of events within the window, and quality indicators; gating weights are applied to the cross-attention fusion results to reduce erroneous associations caused by time mismatches. This gating can affect both spatial location and temporal segment, allowing the model to automatically select the most reliable alignment relationship. To facilitate subsequent S5 alignment based on the "unified geological object primary key index," an "object query token" is introduced: each object query token corresponds to a candidate geological object / region (generated from the AOI grid, buffer, or candidate hazard area); the object query token aggregates spatiotemporal information related to the object from the dual-stream fusion features, outputting object-level spatiotemporal embeddings, naturally adapted to primary key index alignment. The output includes region-level embeddings, as well as object-level and location-level embeddings.

[0081] The improved spatiotemporal Transformer structure achieves the following effects: 1) Enhanced accuracy and robustness of heterogeneous data fusion. The dual-stream structure models the observation mechanisms of remote sensing and InSAR separately. Cross-attention alignment enables the model to establish a correspondence between visible changes in remote sensing and quantized deformation in InSAR. Compared to single-stream or simple stitching, it significantly reduces the probability of "false triggering by changes in remote sensing noise" or "false judgment of low coherence in InSAR" being the real risks. 2) Significantly reduced mismatch caused by time asynchrony and response lag. The time-delay adaptive gating fusion module explicitly considers the observation time difference and the lag relationship between change and deformation, automatically adjusting the alignment weights to reduce the risk of incorrectly associating "irrelevant changes occurring at different times" as the same event, thereby improving the credibility and generalization ability of spatiotemporal embedding. 3) The output is more suitable for subsequent primary key index alignment and evidence chain tracing. The object query token aggregates spatiotemporal information in units of "objects", which enables the embedded representation to be directly bound to the unified geological object primary key index, reducing the alignment cost when constructing cross-modal sample units in S5. At the same time, cross-attention weights naturally provide clues to "which remote sensing change evidence explains which InSAR deformation", enhancing the verifiability and interpretability of subsequent S7 evidence chain construction.

[0082] S5: Construct cross-modal sample units, aligning geological element analysis results, field-level structured records, borehole structured data, and spatiotemporal embedding representations to a unified geological object primary key index, forming traceable training samples;

[0083] In some embodiments, the cross-modal sample unit construction and unified primary key index alignment uses the built-up area and surrounding new areas of a city as the AOI. The system has obtained the lithological filling area mask in S2, the field-level structured records in S3, the borehole structured data in S1, and the spatiotemporal embedding representation of remote sensing video and InSAR time-series video in S4. The goal of this step is to aggregate the above multimodal data onto the same object with the "unified geological object primary key index" as the core, forming traceable training samples for subsequent training of the large geological foundation model in S6.

[0084] The establishment of a unified geological object primary key index (GeoObject Key) in this embodiment uses a composite method of "spatial grid object + business object overlay" to define geological objects: Basic objects: Spatial grid unit objects. The AOI is divided into grid units at a fixed resolution (e.g., 50m×50m or 100m×100m), with each grid unit serving as a candidate geological object. Business objects: Borehole buffer objects (optional enhancement). Buffer zones are constructed centered on borehole locations (e.g., radius 30m–200m, adjusted according to project type) to form borehole-related objects, used to enhance the correspondence between "subsurface stratification and surface changes". Map surface objects: Lithological fill area surface objects (optional enhancement). Connected domains are extracted from the lithological mask output by S2 and vectorized into surface objects (polygons), with each surface object serving as a candidate geological object. The system can uniformly number the above objects to form a primary key index. To ensure traceability and uniqueness, the primary key encoding rule can be generated by combining the following elements: AOI number, object type (mesh / drill buffer / face object), object spatial index (such as mesh row and column number or polygon ID), and version number. For example, the generated primary key would be: geo_obj_id = AOI01-grid-r102-c087-v1 or AOI01-borehole-ZK013-buf100-v1. Simultaneously, the object's geometric boundaries (center point, extent, polygon) are saved for spatial retrieval.

[0085] Cross-modal alignment strategy and sample unit structure. Sample unit data structure (minimum closure of training samples): Each training sample corresponds to a geo_obj_id and contains at least the following four types of information: Geological element analysis results (from S2); Mask patch within the object of the lithological filling area mask, and its position index in the original map (map sheet ID, slice ID, pixel coordinate range); Mask quality markers (porosity, connectivity, boundary complexity, etc., for subsequent sample quality management) can be attached. Field-level structured records (from S3): A set of fields related to the object, such as: project name, report number, borehole number, borehole elevation, top / bottom layer depth, lithological description, etc.; Each field retains the source page ID, text box position, table row and column number, and field confidence (for traceability and weighting). Borehole structured data (from S1 or S3 database results): Borehole location, layer sequence, lithological code, sampling information, etc.; Data source (original database ID / report ID), record version, and update time are retained. The spatiotemporal embedding representation (from S4) is a remote sensing / InSAR spatiotemporal embedding vector within the object's spatial range, or an object-level embedding output from an object query token; it also saves time window information (start and end dates / frame index range) and quality summaries (missing frame rate, average coherence coefficient, etc.). The alignment strategy (spatial + temporal + semantic three-layer alignment) is as follows:

[0086] A) Geological map mask alignment to object primary key (spatial alignment) system maps the spatial boundary of each geo_obj_id to the geological map coordinate system; the mask block within the coverage area of ​​the object is cropped on the mask map as the visual element of the object; if the object spans multiple map sheets or multiple slices, the slice with the largest coverage ratio is selected according to priority, or multiple mask blocks are packaged into a list and the source order is recorded.

[0087] B) Document field records are aligned to the object primary key (semantic + spatial alignment). This embodiment uses a "two-stage binding": First, it is bound to the borehole or project. If a borehole number (such as ZK-013) is identified in the field record, it is directly associated with the borehole_id in the borehole structured data. If no borehole number is identified, it is bound to the project using "project name + report number + page range", and then the spatial location is determined by the known spatial range of the project or the project coordinates. Then, the borehole / project is bound to the geological object primary key to determine which grid / surface object / buffer object the borehole point falls into, that is, the corresponding field record is merged into the geo_obj_id. If the borehole falls at the intersection of multiple candidate objects, it is determined by "closest to the object center" or "largest coverage area".

[0088] C) Align borehole structured data to object primary keys (spatial alignment + consistency check) Quickly retrieve the object to which the borehole point belongs through spatial index; use the layer sequence and lithological properties of the borehole as the underground structural features of the object; if there are multiple boreholes within the object range, form a set according to rules: for example, sort by distance and take the top N, or perform statistical aggregation (layer thickness distribution, lithological ratio, etc.).

[0089] D) Spatiotemporal embedding alignment to object primary key (spatial alignment + time window alignment): For each object, the system extracts embeddings that overlap with the spatial extent of the object from the S4 output: if the S4 output is a grid / patch embedding, the set of patches covered by the object is selected and aggregated into an object embedding; if the S4 output contains an object query token, the object embedding corresponding to that token is directly retrieved. Simultaneously, the embedding is bound to a specified time window (e.g., the last 30 days, the last 12 InSAR observations), and the window parameters and missing data are recorded.

[0090] To ensure traceability, this embodiment writes the following traceability information to each training sample: Primary key level traceability: geo_obj_id, object geometric boundary, generation rule version number. Map mask traceability: map sheet ID, original image file hash, slice ID, cropping pixel coordinate range, mask threshold, and post-processing parameter version. Document field traceability: report ID, page ID, field source text box position (rectangular or polygonal coordinates), table row and column numbers, OCR confidence, and field mapping rule version number. Borehole data traceability: original data source identifier (database / report), borehole ID, record version number, and update time. Video embedding traceability: remote sensing video clip ID and timestamp range, InSAR frame index and date range, quality summary metrics (coherence, missing frame rate), and embedding model version number.

[0091] Sample quality control and inclusion strategy (avoiding noise contamination in training): To improve the reliability of training samples, the system performs quality assessment on sample units and determines the inclusion method: Prerequisites: The object must possess at least one of "spatiotemporal embedding" or "geological map mask" and have a locatable primary key; Confidence threshold: If the OCR confidence of a field record is too low, it is marked as weakly supervised or excluded from key field training; Consistency verification: If there is a significant semantic conflict between borehole stratification and lithology mask (e.g., lithology code and legend mapping do not match and cannot be explained), the sample enters the "waiting-for-verification queue" and is not included in the high-confidence training set; Sample stratification: Samples are divided into three levels based on quality: high / medium / low. High-quality samples are used for supervised training, while medium and low-quality samples are used for self-supervised or comparative learning tasks.

[0092] The final output of the example system is a set of "cross-modal sample units". Each sample unit is uniquely indexed by geo_obj_id and contains: lithological filling area mask blocks and their source locations; field-level structured record sets and their document traceability information; borehole structured layered sequences and their source information; remote sensing and InSAR object-level / location-level spatiotemporal embedding and time window traceability information; thus forming traceable training samples that can be directly used to train large-scale geological models in S6.

[0093] S6: Train a large geological foundation model based on traceable training samples. The large geological foundation model includes a multimodal encoder, a cross-modal fusion backbone network, and a task output head. The multimodal encoder includes a video encoder.

[0094] In some embodiments, a large-scale geological model is trained based on traceable training samples. The implementation scenarios and training data sources are constructed in S5 using cross-modal sample units. Each sample unit uses a unified geological object primary key index, geo_obj_id, as its unique key and includes: geological element analysis results: lithological filling area masks (and their corresponding map clipping blocks and source locations); field-level structured records: OCR fields from scanned images of exploration documents (including source page, text box position, table rows and columns, and field confidence); borehole structured data: borehole locations, layer sequences, lithological attributes, etc.; spatiotemporal embedding representations or traceable remote sensing video clips and InSAR time-series video windows (including clip identifiers, timestamps / date ranges, and quality information). The system aggregates all samples by object primary key and divides training samples by time window (e.g., "Recent N periods of InSAR + corresponding time period remote sensing changes" as a window sample) to form training, validation, and test sets. Simultaneously, sample traceability information (source file hash, page ID, video clip ID, coordinate range, etc.) is bound to the model training log to ensure the training process is verifiable.

[0095] Training Process (Phased Training to Ensure Stable Convergence and Generalization) This embodiment adopts a "two-stage training + multi-task joint optimization" approach: Stage A: Cross-modal alignment pre-training (weakly supervised / self-supervised) The purpose is to allow the model to first learn the "consistency of different modalities under the same primary key object", reducing the dependence of subsequent task training on annotations. Typical training tasks include: Video-image consistency comparison: bringing video spatiotemporal segments with the same geo_obj_id closer to the lithological mask clipping block; and widening the representation of samples from different objects; Video-bore consistency comparison: aligning video spatiotemporal segments with borehole layer sequence representations; Field-bore consistency verification: aligning "layer depth / lithological description" and other fields in the OCR field with the borehole structured fields; Temporal mask reconstruction (optional): randomly occluding several time slices in the InSAR temporal window, requiring the model to recover trend features from the context, thereby enhancing robustness to missing data and noise.

[0096] Phase B: Joint Training of Target Tasks (Supervised / Weakly Supervised Hybrid) Joint training is performed by overlaying the business task output heads on top of Phase A. Optional target tasks include: Deformation Anomaly Identification: Outputting the anomaly level or risk score of the object within a given time window; Geological Attribute Inference: Outputting the stratigraphic unit category, lithological code, or engineering geological zoning category corresponding to the object; Cross-Modal Consistency Scoring: Outputting the consistency confidence between "map lithological mask—borehole stratification—spatiotemporal deformation," used for quality control and sample selection. To utilize the quality information of "traceable training samples," this embodiment sets quality weights for training samples: when OCR confidence is low, InSAR coherence is low, or mask quality is poor, the sample's contribution to the loss decreases, thereby reducing the contamination of the large model by noisy samples.

[0097] The geological foundation large model includes the following components: 1) a multimodal encoder (including at least a video encoder); 2) a cross-modal fusion backbone network; and 3) a task output head.

[0098] The components and functions of a multimodal encoder are as follows:

[0099] The video encoder takes as input remote sensing video segments (continuous frames) and their deformation proxy channels (e.g., inter-frame displacement / intensity of change), or object-level / location-level spatiotemporal embeddings from S4 output; InSAR temporal video windows (deformation frame sequences arranged by date) and quality channels (coherence coefficient / missing markers). Frame-level feature extraction extracts spatial features from each frame, preserving local texture and regional change clues; temporal aggregation encoding encodes the frame sequence in the temporal dimension, learning trends, abrupt changes, periodicity, and hysteresis responses; outputs "object-level video embeddings" (spatiotemporal vectors corresponding to a certain geo_obj_id) and / or "location-level video embeddings" (vector sets corresponding to the internal grid / patch of the object). The output video modal embedding set (with temporal window identifiers and quality summaries) serves as one of the main inputs to the fusion backbone.

[0100] The image / mask encoder (for lithology-filled area masks) takes as input geological map clipping blocks and corresponding lithology-filled area masks (which can be used as additional channel overlay inputs). It extracts lithological spatial distribution morphology, boundary complexity, texture distribution, etc., to form "map semantic embeddings." The output is a map embedding vector or a patch-level embedding set.

[0101] The text field encoder (for field-level structured records) takes in field-level structured records (field code, field value, unit, confidence level, field source type, etc.). Field serialization organizes fields into a sequence according to the template order; semantic encoding encodes field names and values ​​into fusionable vector representations; location and source encoding uses the field's source page / table location as an auxiliary marker to enhance traceability. The output is a sequence of text field embeddings and aggregated object-level text embeddings.

[0102] The structured data encoder (for borehole stratigraphic sequences) takes as input the stratigraphic sequence (top / bottom depth, lithological code / description, age / stratigraphic unit, etc.) from the borehole structured data, as well as the spatial location of the boreholes and their association with objects. This function encodes the "stratigraphic sequence" into a sequence representation that can be aligned with video / maps, highlighting geological sequences and lithological assemblage patterns. Outputs borehole embedding sequences and object-level aggregate embeddings; when multiple boreholes exist within an object, it outputs a "multi-borehole set embedding." Implementations B / C / D are preferred and can be written in the specification using the "may include" clause, without locking the claims; however, from the perspective of system integrity and implementability, it is recommended that you explicitly use at least two of them in the specification.

[0103] The cross-modal fusion backbone network structure and function are as follows: The input fusion backbone receives embedding sets from each encoder, including at least: object-level / position-level embeddings from the video encoder output; and one or more of image mask embeddings, text field embeddings, and drill embeddings. All embeddings are identified by geo_obj_id for easy aggregation at the object level.

[0104] This embodiment employs an "object query-driven cross-attention fusion" structure: One or more "object query vectors" are set for each geo_obj_id as the fusion entry point; the object query vectors retrieve and aggregate the most relevant information from video embeddings, image embeddings, field embeddings, and borehole embeddings; a gating mechanism is introduced during the aggregation process: when a modality has low quality (e.g., low coherence in InSAR, low confidence in OCR), its contribution is automatically reduced to prevent noise from dominating the fusion result; after fusion, a "unified object representation vector" is obtained, serving as the shared input for the downstream task output header. The output is a unified object representation (object-level fusion embedding); optional outputs include cross-modal attention weights and modal contribution weights (used for subsequent S7 evidence chain construction and audit review).

[0105] The specific functions and output formats of the task output heads are as follows. The task output heads take a "unified object representation vector" as input and can be configured with one or more output heads to support multi-task training and applications: Deformation / Risk Output Head: Outputs the object's risk level, anomaly score, or trend category (e.g., stable / gradual / accelerated / abrupt). Uses: Monitoring and early warning, hazard screening, risk ranking. Geological Attribute Output Head: Outputs stratigraphic unit category, lithological category / code, engineering geological zoning category, etc. Uses: Urban geological mapping assistance, geological model updating, planning support. Consistency and Quality Output Head: Outputs cross-modal consistency score, sample credibility score, key field reliability score. Uses: Training sample screening, low-quality data re-verification, system self-checking. Uncertainty Output: The confidence level or confidence interval of each task result, used to distinguish between "model-confident conclusions" and "conclusions requiring manual verification."

[0106] After training is completed, the S6 training output and delivery system outputs: parameters of the trained geological foundation model (including version number); model version information of each modal encoder and fusion backbone; traceability logs bound to each training batch (sample primary key, source file / page / video clip identifier, quality weight) for auditing and experiment reproduction; and intermediate products for S7: object-level fusion embedding, cross-modal attention weights, and key time segment contribution ranking, etc.

[0107] S7: The geological foundation large model outputs the target task results and the evidence chain corresponding to the target task results. The evidence chain includes the primary key index, the source video segment identifier and timestamp, the source image region location, and the source document paragraph location.

[0108] In some embodiments, taking urban geological monitoring and early warning as an example, the system outputs the target task results and generates the evidence chain for the target task, using the unified geological object primary key index geo_obj_id as the unit. The target task can be one or more of the following (this embodiment focuses on "deformation anomaly identification," but is compatible with other tasks): Deformation anomaly identification: outputs the anomaly level and risk score of the object within a specified time window; Trend type determination: outputs trend categories such as stable, gradual change, acceleration, and abrupt change; Geological attribute inference: outputs lithological category, stratigraphic unit, or engineering geological zoning category. In addition to the result itself, the system simultaneously outputs the evidence chain corresponding to the result for review and auditing by business personnel. Input conditions (from previous steps): For each geo_obj_id, the system already has: Video-side input: encoded representations of remote sensing video clips and InSAR time-series video windows, with each clip retaining a clip identifier, start and end timestamps, or date range; Image-side input: geological map clipping blocks and lithological fill area masks, retaining map sheet ID, slice ID, and pixel coordinate range; Document-side input: field-level structured records and their source page IDs, paragraph / table row and column positions, and text box coordinate ranges; Alignment index: the association between the object primary key index geo_obj_id and the above multimodal source data (from S5). Target task result output method (object-level output): The system outputs one object-level result record for each geo_obj_id, including at least: geo_obj_id (primary key index); target task result (e.g., anomaly level, risk score, trend category, or geological attribute category); result confidence level; and the time window identifier corresponding to the result (e.g., start and end dates or frame index range). This result record serves as the "result end" of the evidence chain, used for association with the evidence ends.

[0109] This embodiment uses a "contribution-driven evidence retrieval and localization" approach to generate the evidence chain. The core idea is to extract the "input segment / region / field that contributes most to the result" from the model's cross-modal fusion process and task output header, and map it back to the locatable coordinates of the original data. The video evidence localization (source video segment identifier and timestamp) candidate video segment set determination system retrieves the set of remote sensing video segments and the InSAR time-series video window set (including segment ID, timestamp / date range, and quality summary) related to the object from the S5 association table based on geo_obj_id and the time window. The contributing segment selection system identifies several video segments with the "highest contribution" from the modal contribution weights or attention weights output by the fusion backbone; if an object query vector is used, the time segment and frame segment with the highest contribution to the query for that object are selected. For each selected segment, the system outputs the following: a video segment identifier (e.g., rs_clip_id or insar_window_id); the corresponding timestamp range (start and end times) or observation date range; and, if it is an InSAR sequence, the system also outputs the keyframe index in the "frame index-date" mapping for quick playback and location (optional enhancement). Video evidence entries are generated and written into the evidence chain, serving as a "playbackable and verifiable" evidence entry point.

[0110] The image evidence localization (source image region location) candidate image region determination system obtains the associated geological map slice information (map sheet ID, slice ID) and the spatial projection range of the object in the map based on geo_obj_id. Key region selection is based on the model's contribution weight in image mask embedding, or on mask boundary complexity and overlap with anomalous regions, selecting the image region that contributes the most to the result. Location representation outputs for each key region: source image identifier (map sheet ID / slice ID); image region location (preferably pixel coordinate range or polygon coordinates); and the local cropping ID of the corresponding lithological filling area mask, used for one-click overlay display. Image evidence entry generation writes the above location fields into the evidence chain, enabling auditors to directly locate the "lithological distribution / interface region that generates the judgment basis" on the map.

[0111] The document evidence location (source document paragraph position) candidate field record determination system retrieves a set of field-level structured records associated with the object geo_obj_id. Each record includes: report ID, page ID, paragraph / table row and column, text box coordinates, and field confidence score. Key field selection filters fields with high contributions from the model fusion weights, preferably including: stratification depth, lithological description, borehole elevation, groundwater level, and test indicators—fields related to the target task. If the target task is risk identification, fields related to foundation stability, groundwater, and soft soil layers are prioritized. Paragraph position expression outputs for each key field: source document identifier (report number / report ID); source page ID and page number; paragraph position (paragraph number in the main text) or table position (row number, column number); and text box coordinate range (for easy highlighting on the PDF rendering). Document evidence item generation writes the above fields into the evidence chain, allowing reviewers to directly jump to specific pages, paragraphs, or cells to verify the original description.

[0112] For each geo_obj_id, the system outputs a "Result Record + Evidence Chain Record". The evidence chain record must include at least: Primary Key Index: geo_obj_id; Video Evidence Set: Several entries, each containing "Source Video Segment Identifier, Timestamp / Date Range (and optional Keyframe Index), Quality Summary"; Image Evidence Set: Several entries, each containing "Source Map Sheet / Slice Identifier, Image Region Location (Pixel Coordinate Range / Polygon Coordinates), Mask Cropping Identifier"; Document Evidence Set: Several entries, each containing "Report ID, Page ID / Page Number, Paragraph Number or Table Row / Column Number, Text Box Coordinate Range, Field Code and Field Value Summary". Each evidence entry in the evidence chain can be used to retrieve the original data objects from S1 to S5, achieving a closed-loop traceability "from result to original material".

[0113] For example, when geo_obj_id=AOI01-grid-r102-c087-v1 is determined to have a "high risk of accelerated subsidence", the system outputs the following evidence chain: Video evidence: InSAR window insar_window_2024Q3, with the key time period being "July-September", pointing to several observation dates that contributed the most; at the same time, the timestamp range of the remote sensing clip rs_clip_0831_roadA is given for playback to view signs of surface changes; Image evidence: A region coordinate range from the geological map map_1_5000_2022_tile_34, corresponding to an area with complex lithological filling boundaries and consistent with the distribution of weak soil layers; Document evidence: The field "lithological description = soft plastic silty clay" in the 5th row and 3rd column of the table on page 12 of the survey report report_2023_1120, and the field "groundwater level = shallow" on the same page, used to explain the source of subsidence sensitivity. Based on this, reviewers can directly open the corresponding video clip, locate the image area, and jump to the report page to verify the original text of the fields.

[0114] Preferably, the U-Net for semantic segmentation of the lithological filling area in step S2 is an improved U-Net, which includes a multi-scale context encoding branch and a boundary refinement branch; wherein, the multi-scale context encoding branch extracts lithological texture features of different receptive fields through dilated convolution, and the boundary refinement branch enhances the boundary of the lithological filling area through an edge-guided attention module to reduce omissions and misclassifications at geological interfaces.

[0115] Preferably, the improved U-Net sets a cross-layer feature alignment gating module at the skip connection. The cross-layer feature alignment gating module generates gating weights based on the correlation between the features of the encoding end and the decoding end, and adaptively fuses the low-level texture features and high-level semantic features, thereby improving the segmentation robustness of the lithological filling area under the interference of complex symbols and annotations.

[0116] Preferably, step S2 further includes performing layout structure analysis on the geological map raster image, identifying the lithological filling area, legend area, annotation area and border area, and masking the legend area and annotation area to reduce the interference of non-lithological areas on the generation of the lithological filling area mask.

[0117] Preferably, the field structure extraction in step S3 includes table structure recovery. The table structure recovery obtains the table topology by detecting cell boundaries and inferring row and column relationships, and matches the cell content with the field template to generate borehole number, coordinates, elevation, layer depth, and lithological description.

[0118] Preferably, the spatiotemporal Transformer in step S4 adopts a dual-stream coding structure, including a remote sensing video change coding stream and an InSAR temporal deformation coding stream; wherein, the change coding stream encodes the temporal change features of the remote sensing video, the deformation coding stream encodes the deformation sequence of the InSAR temporal video, and the change features and deformation features are aligned and fused through cross-attention modules to generate the spatiotemporal embedding representation.

[0119] Preferably, the alignment to the unified geological object primary key index in step S5 includes: establishing a correlation between the spatial range of the video clip and the geological map raster image, and between the borehole location and the InSAR location based on a joint matching strategy of spatial neighborhood and time window; wherein the spatial neighborhood is determined by the buffer radius, and the time window is determined by the sampling interval or the duration of the event.

[0120] Preferably, the multimodal encoder in step S6 includes a structured data encoder, which encodes the layered sequence and lithological properties in the borehole structured data into a sequence embedding, and aligns it with the spatiotemporal embedding representation output by the video encoder through a cross-modal fusion backbone network to learn the joint representation between the lithological filling mask-borehole layer-deformation sequence.

[0121] Preferably, the evidence chain output in step S7 further includes uncertainty information and counterfactual explanation information. The uncertainty information includes confidence scores or confidence intervals, and the counterfactual explanation information includes key frame indexes, key image region masks and their corresponding document field references that contribute the most to the target task results, thereby supporting the auditing and review of the output results.

[0122] This application also provides a geological large model construction system based on video processing, such as Figure 2The device includes the following hardware units: Remote sensing video acquisition equipment: one or more of the following: UAV-borne camera / flying camera, ground patrol camera, and vehicle-mounted camera, used to acquire remote sensing video and generate video files and timestamp information; InSAR data access equipment / interface: a server interface that interfaces with the InSAR processing platform or monitoring platform, used to access InSAR time-series deformation raster sequences and their observation date information, and can encapsulate them into InSAR time-series video; Document scanning equipment: one or more of the following: a scanner / high-speed document scanner / multifunction printer, used to acquire scanned images or PDFs of survey reports, borehole columnar sections, test tables, etc.; Borehole data acquisition terminal: a field recording terminal / database import interface, used to import borehole structured data. The computing server cluster (core) consists of application server nodes (primarily CPU): used for running general logic such as data access, task scheduling, index alignment, dictionary mapping, evidence chain organization, and interface services; inference server nodes (primarily GPU): used for running geological map U-Net segmentation inference, text detection and recognition inference, spatiotemporal Transformer inference, and cross-modal fusion inference; training server nodes (primarily GPU, which can be reused with inference nodes or used independently): used for training large geological foundation models, supporting multi-GPU parallel training and model version management; and load balancing and message queue nodes (optional): used for high-concurrency task distribution, asynchronous processing, and failure retries. Storage and Database Subsystem: Object Storage / File Storage: Saves raw video, InSAR frame sequences, scanned documents, geological map raster, model files, slices, and intermediate products; Relational Database: Saves field-level structured records, borehole structured data, data dictionary, task status, model version, sample index, etc.; Spatial Database (GIS Database): Saves geometric boundaries of geological objects, spatial indexes, vectorized surface objects, borehole locations, AOI boundaries, etc.; Log and Audit Storage: Saves training traceability information, evidence chain index, access logs, and operation logs. User Interaction and Visualization Terminal Client / Workstation: Used for data uploading, task configuration, result viewing, and evidence chain verification; Large Screen / Monitoring Terminal (Optional): Used for risk ranking, time-series trend display, and early warning issuance. Communication Network Wired / Wireless Network: Used for data transmission between the acquisition end and the server, and between the server and the storage; security can be ensured using LAN / private network / VPN, etc.

[0123] The data path connects remote sensing video acquisition equipment to upload video files to object storage via wireless links or storage media; simultaneously, it writes the acquisition timestamp and trajectory / attitude (if any) to a relational database through an interface. The InSAR access interface server pulls deformation raster sequences and observation date tables from an external InSAR platform, writes the sequences to object storage, and writes quality information such as "frame index-date mapping and coherence coefficient" to a relational / spatial database. The document scanning equipment uploads PDF / page images to object storage and writes metadata such as report number, page number, and project name to the relational database. The borehole data acquisition terminal / import interface writes structured tables such as borehole locations and layer sequences to the relational database, and its spatial fields (coordinate points) are simultaneously written to the spatial database.

[0124] The computational pathway connects (modules deployed to hardware). Application server nodes pull pending tasks and data lists from the database and distribute inference tasks to GPU inference server nodes via message queues (optional). GPU inference server nodes read map tiles, document pages, remote sensing video clips, and InSAR frame sequences from object storage and perform: U-Net geological element visual analysis inference; text detection + text recognition cascaded OCR inference; and spatiotemporal Transformer encoding inference. The inference output masks, text boxes, recognition results, and spatiotemporal embeddings are written back to object storage and the database. Application server nodes utilize the spatial indexing capabilities of the spatial database to align mask results, field records, borehole data, and spatiotemporal embeddings along the object primary key dimension, generating a traceable training sample index and writing it to the database. Training server nodes read traceable training samples from object storage / database, perform geological foundation large-scale model training, and write model weights, configurations, and version information to object storage and the database after training. They can also hot-update the inference service to the GPU inference nodes. The output and evidence chain service is implemented by the application server: It outputs the target task results for each geological object's primary key, and simultaneously aggregates evidence chains from the database, including "video clip ID and timestamp, image region coordinates, and document paragraph / table position," and returns them to the client. 3) User review and visualization connection: The client / workstation connects to the application server via HTTPS / private network interface to obtain results and evidence chains; the client initiates a video playback request to the object storage based on the "video clip identifier and timestamp" in the evidence chain; it loads the corresponding map slice and overlays a mask based on the "image region position"; and it opens the PDF based on the "document paragraph position" and highlights the positioned text box / table cell, enabling one-click review.

[0125] Data Access Module: Deployed on the application server, connecting object storage, relational databases, and spatial databases; Geological Feature Visual Analysis Module (U-Net): Deployed on the GPU inference server, reading map tiles and outputting lithological masks; Field-Level Structured Record Extraction Module (Detection + Recognition Cascade): Deployed on the GPU inference server, reading document pages and outputting field records and coordinates; Spatiotemporal Embedding Representation Generation Module (Spatiotemporal Transformer): Deployed on the GPU inference server, reading remote sensing / InSAR video windows and outputting spatiotemporal embeddings; Cross-Modal Fusion Module: Logically, the application server completes primary key alignment and sample construction, while model-level fusion is completed on the GPU inference / training nodes; Training Module: Deployed on the training server (GPU), reading traceable training samples and training a large-scale geological foundation model; Output Module: Deployed on the application server, outputting task results and evidence chains, and providing a verification interface.

[0126] The data access module acquires a set of urban geological data, which includes: geological map raster images, scanned images of exploration documents, remote sensing videos, InSAR time-series videos, and borehole structured data.

[0127] The geological element visual analysis module performs geological element visual analysis on the geological map raster image using U-Net with semantic segmentation of lithological filling area. The output geological element analysis results include: lithological filling area mask.

[0128] The field-level structured record extraction module performs image text recognition and field structure extraction on the scanned image of the exploration document using a cascaded model of a text detection network and a text recognition network, outputs field-level structured records, and maps the field-level structured records to a preset geological data dictionary;

[0129] The spatiotemporal embedding representation generation module uses a spatiotemporal Transformer to extract spatiotemporal features from the deformation sequence of the remote sensing video and InSAR time-series video and encodes them to generate spatiotemporal embedding representations.

[0130] The cross-modal fusion module constructs cross-modal sample units, aligning geological element analysis results, field-level structured records, borehole structured data, and spatiotemporal embedding representations to a unified geological object primary key index, forming traceable training samples;

[0131] The training module trains a large geological foundation model based on traceable training samples. The large geological foundation model includes a multimodal encoder, a cross-modal fusion backbone network, and a task output head, wherein the multimodal encoder includes a video encoder.

[0132] The output module outputs the target task results from the geological foundation large model, and outputs the evidence chain corresponding to the target task results. The evidence chain includes the primary key index, the source video segment identifier and timestamp, the source image region location, and the source document paragraph location.

[0133] Those skilled in the art will understand that embodiments of this application may be provided as methods, systems, or computer program products, and therefore this application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0134] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A method for constructing a large geological model based on video processing, characterized in that, Including the following steps: S1: Obtain a set of urban geological data, which includes: geological map raster images, scanned images of exploration documents, remote sensing videos, InSAR time-series videos, and borehole structured data; S2: Perform geological feature visual analysis on the geological map raster image using U-Net with semantic segmentation of lithological filling area. The output geological feature analysis results include: lithological filling area mask; S3: The cascaded model of text detection network and text recognition network is used to perform image text recognition and field structure extraction on the scanned image of the exploration document, outputting field-level structured records, and mapping the field-level structured records to a preset geological data dictionary; S4: The remote sensing video and InSAR time series video are processed by a spatiotemporal Transformer to extract spatiotemporal features from the deformation sequence and encode them to generate a spatiotemporal embedding representation; S5: Construct cross-modal sample units, aligning geological element analysis results, field-level structured records, borehole structured data, and spatiotemporal embedding representations to a unified geological object primary key index, forming traceable training samples; S6: Train a large geological foundation model based on traceable training samples. The large geological foundation model includes a multimodal encoder, a cross-modal fusion backbone network, and a task output head, wherein the multimodal encoder includes a video encoder. S7: The geological foundation large model outputs the target task results and the evidence chain corresponding to the target task results. The evidence chain includes the primary key index, the source video segment identifier and timestamp, the source image region location, and the source document paragraph location.

2. The method for constructing a large geological model based on video processing according to claim 1, characterized in that, The U-Net for semantic segmentation of the lithological filling area in step S2 is an improved U-Net, which includes a multi-scale context encoding branch and a boundary refinement branch. The multi-scale context encoding branch extracts lithological texture features from different receptive fields through dilated convolution, and the boundary refinement branch enhances the boundary of the lithological filling area through an edge-guided attention module to reduce omissions and misclassifications at geological interfaces.

3. The method for constructing a large geological model based on video processing according to claim 2, characterized in that, The improved U-Net sets a cross-layer feature alignment gating module at the skip connection. The cross-layer feature alignment gating module generates gating weights based on the correlation between the features of the encoder and decoder, and adaptively fuses the low-level texture features and high-level semantic features, thereby improving the segmentation robustness of the lithological filling area under the interference of complex symbols and annotations.

4. The method for constructing a large geological model based on video processing according to claim 1, characterized in that, Step S2 further includes performing layout structure analysis on the geological map raster image, identifying the lithological filling area, legend area, annotation area and border area, and masking the legend area and annotation area to reduce the interference of non-lithological areas on the generation of the lithological filling area mask.

5. The method for constructing a large geological model based on video processing according to claim 1, characterized in that, The field structure extraction in step S3 includes table structure recovery. The table structure recovery obtains the table topology by detecting cell boundaries and inferring row and column relationships, and matches the cell content with the field template to generate borehole number, coordinates, elevation, layer depth, and lithology description.

6. The method for constructing a large geological model based on video processing according to claim 1, characterized in that, The spatiotemporal Transformer in step S4 adopts a dual-stream coding structure, including a remote sensing video change coding stream and an InSAR temporal deformation coding stream. The change coding stream encodes the temporal change features of the remote sensing video, and the deformation coding stream encodes the deformation sequence of the InSAR temporal video. The change features and deformation features are aligned and fused through cross-attention modules to generate the spatiotemporal embedding representation.

7. The method for constructing a large geological model based on video processing according to claim 1, characterized in that, The alignment to the unified geological object primary key index in step S5 includes: establishing a connection between the spatial range of video clips and geological map raster images, and between borehole locations and InSAR locations based on a joint matching strategy of spatial neighborhood and time window; wherein the spatial neighborhood is determined by the buffer radius, and the time window is determined by the sampling interval or event duration.

8. The method for constructing a large geological model based on video processing according to claim 1, characterized in that, The multimodal encoder in step S6 includes a structured data encoder, which encodes the layered sequence and lithological properties in the borehole structured data into sequence embeddings and aligns them with the spatiotemporal embedding representation output by the video encoder through a cross-modal fusion backbone network to learn the joint representation between the lithological filling mask, borehole layering, and deformation sequence.

9. The method for constructing a large geological model based on video processing according to claim 1, characterized in that, The evidence chain output in step S7 also includes uncertainty information and counterfactual explanation information. The uncertainty information includes confidence scores or confidence intervals, and the counterfactual explanation information includes key frame indexes, key image region masks and their corresponding document field references that contribute the most to the target task results, thereby supporting the auditing and review of the output results.

10. A geological large-scale model construction system based on video processing, characterized in that, include: The data access module acquires a set of urban geological data, which includes: geological map raster images, scanned images of exploration documents, remote sensing videos, InSAR time-series videos, and borehole structured data. The geological element visual analysis module performs geological element visual analysis on the geological map raster image using U-Net with semantic segmentation of lithological filling area. The output geological element analysis results include: lithological filling area mask. The field-level structured record extraction module performs image text recognition and field structure extraction on the scanned image of the exploration document using a cascaded model of a text detection network and a text recognition network, outputs field-level structured records, and maps the field-level structured records to a preset geological data dictionary; The spatiotemporal embedding representation generation module uses a spatiotemporal Transformer to extract spatiotemporal features from the deformation sequence of the remote sensing video and InSAR time-series video and encodes them to generate spatiotemporal embedding representations. The cross-modal fusion module constructs cross-modal sample units, aligning geological element analysis results, field-level structured records, borehole structured data, and spatiotemporal embedding representations to a unified geological object primary key index, forming traceable training samples; The training module trains a large geological foundation model based on traceable training samples. The large geological foundation model includes a multimodal encoder, a cross-modal fusion backbone network, and a task output head, wherein the multimodal encoder includes a video encoder. The output module outputs the target task results from the geological foundation large model, and outputs the evidence chain corresponding to the target task results. The evidence chain includes the primary key index, the source video segment identifier and timestamp, the source image region location, and the source document paragraph location.

Images