A two-dimensional and three-dimensional seismic geological feature interactive identification method based on task separation lightweight adaptation and agent routing

Abstract

The present application relates to a kind of two-dimensional and three-dimensional seismic geologic feature interactive identification method based on task separation light adaptation and agent routing.It relates to the field of geophysical exploration technology, and the method comprises the following steps:1, two-dimensional slice and three-dimensional seismic body are constructed into video target segmentation sequence and are differentially pretreated;2, seven types of geological targets are respectively established independent training task and are configured LoRA low rank adaptation branch;3, training and saving include LoRA matrix and prompt, mask and memory module training parameter target special adaptation parameter set;4, model registry and constrained agent routing are established;5, two-dimensional interactive segmentation, three-dimensional bidirectional propagation and memory backwrite error correction are executed.The method can reduce the risk of manual selection error and direct control model state of generative output;Interpretation personnel can repair three-dimensional continuity with a small amount of interaction, avoid slice-by-slice depiction;Error correction process and model version can be recorded, which is convenient for review.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration technology, specifically to an interactive identification method for seismic geological targets such as faults, stratigraphic positions, river channels, seismic facies, karst caves, chimneys, and salt bodies. This method comprises a closed loop consisting of a target-specific adaptation parameter set, a model registry, constrained agent routing, deterministic parameter loading, and two-dimensional / three-dimensional prompting segmentation and memory error correction. Background Technology

[0002] Seismic exploration utilizes the reflection response of artificially induced seismic waves at subsurface interfaces to characterize stratigraphic structure and lithological variations. Identifying geological targets such as faults, stratigraphic horizons, channels, seismic facies, caves, chimneys, and salt bodies is a crucial foundation for tectonic interpretation, sedimentary system analysis, reservoir prediction, trap evaluation, and well placement. With the continuous increase in the spatial extent and sampling density of 3D seismic data, interpreters need to repeatedly locate, delineate, and verify the same targets on numerous master lines, connecting lines, and time or depth slices. Slice-by-slice manual interpretation is time-consuming, and different interpreters may have inconsistent judgments regarding weak reflections, ambiguous boundaries, and complex overlay relationships. Therefore, how to reduce redundant depiction while preserving the interactive control capabilities of interpreters has become a fundamental problem that needs to be solved in intelligent seismic data interpretation.

[0003] Unlike objects in natural images that have clear outlines and stable appearances, seismic geological targets are typically represented indirectly through the discontinuity, termination, bending, amplitude anomalies, texture changes, and spatial relationships of their reflection phase axes. Significant scale and morphological differences also exist between different targets: faults often appear as discontinuous zones and phase axis breaks; stratigraphic layers appear as slender, laterally extending interfaces; river channels and salt bodies usually have continuous regional boundaries; caves often exhibit localized strong and weak amplitude anomalies; and chimneys show vertical disturbances or energy attenuation. The same local reflection anomaly may correspond to different geological origins, and the same target may present different appearances in different work areas, with different slice orientations, or under different data processing procedures. This characteristic of "similar local responses but different geological semantics, and unstable appearances of similar targets" means that multi-target recognition cannot be simply viewed as natural image category segmentation.

[0004] Existing automatic seismic target identification methods mainly include thresholding or rule-based methods based on attributes such as coherence, variance, curvature, and texture, as well as supervised segmentation methods based on models such as convolutional neural networks and Transformers. Attribute-based methods rely on manually selecting attribute combinations and thresholds, and complex structures, strong noise, and similar anomalies can easily lead to false detections. Supervised segmentation models can learn the features of specific targets, but they usually perform fixed inference according to predetermined input sizes, categories, and output channels, making it difficult to quickly assign new target instances through minimal interaction after training. For 3D seismic volumes, directly using 3D networks also incurs high memory and computational overhead; while using 2D networks for slice-by-slice prediction has lower overhead, it lacks consistency constraints across slices, easily leading to breaks, jumps, and instance confusion between adjacent slices.

[0005] To reduce manual annotation and enhance interactivity, general cue segmentation models can utilize positive points, negative points, bounding boxes, or coarse masks to generate target masks without pre-setting specific instance locations. These models offer the possibility of "interpreters providing minimal spatial constraints, and the model completing boundary expansion" for seismic interpretation. However, general cue segmentation models are typically trained on natural images, whose feature representations are biased towards object contours, colors, and common textures; seismic images, on the other hand, primarily consist of grayscale amplitudes and continuous reflection structures, and target boundaries are often not closed. Models not adapted to the seismic domain tend to include reflection areas adjacent to cue points but with different geological significance as targets, or only extract local strong responses while omitting continuously extending portions. The sparser the cue, the more the model relies on its internal semantic representation, and the more pronounced the target leakage and omission caused by the aforementioned domain differences become.

[0006] Furthermore, video target segmentation models with memory mechanisms can encode keyframe cues and masks into memories and propagate the results to adjacent frames. This mechanism can be used to reduce slice-by-slice drawing if a 3D seismic body is unfolded into an ordered sequence of slices along a certain direction. However, 3D geological bodies may bifurcate, merge, pinch out, experience abrupt boundary changes, or suffer from reduced reflection quality along the slice direction; the movement between adjacent slices is not simply a rigid target motion. When the initial mask of a key slice is incomplete, or when propagation passes through slices with low signal-to-noise ratios or abrupt target morphology changes, errors are written into subsequent memories and continue to propagate. Simply replacing a 2D mask on an error slice without re-entering the correction result as a conditional memory and influencing the propagation process on both sides of that slice cannot restore the overall continuity of the 3D target.

[0007] To reduce the differences between the natural image domain and the seismic image domain, full-parameter fine-tuning, joint multi-class fine-tuning, or low-rank parameter fitting can be applied to the pre-trained model. Full-parameter fine-tuning requires updating and storing a large number of model parameters, which is not conducive to the independent maintenance and incremental expansion of multi-objective models. Low-rank parameter fitting can adjust the attention projection with fewer new parameters while freezing the pre-training backbone, which is a way to reduce training and storage costs. However, if targets with significant morphological differences, such as faults, stratigraphy, and river channels, share the same set of trainable and updatable parameters, the gradients in joint training may compete with each other, making it difficult for the model to determine whether the user is referring to a slender interface, a regional geological body, or an anomalous texture under weak cue conditions. Therefore, forming separate fitting parameters for geological targets can reduce interference between targets, but it also changes the model management problem from "one model handling multiple categories" to "one shared base corresponding to multiple sets of target-specific fitting parameter sets".

[0008] When multiple sets of target-specific adaptation parameter sets coexist, online identification requires not only knowing the target name but also simultaneously determining the model configuration, adaptation weights, backbone structure, input resolution, and availability status that match the target. Geological targets may also be described using abbreviations, aliases, or a mixture of Chinese and English in actual instructions. If interpreters manually search for configuration and weight files, problems such as inconsistent understanding of aliases, mismatched configuration and weight versions, mismatched backbone structures, or duplicate injection of the same parameters can easily occur. Simply saving multiple checkpoint files cannot establish a stable correspondence between natural language targets, specification categories, and actual loaded parameters; therefore, a searchable and verifiable model organization method is needed.

[0009] Language models possess natural language understanding capabilities, extracting geological targets, data paths, and operational intentions from user commands, thus reducing the burden on interpreters to remember model names and weight paths. However, language model outputs are generative and uncertain. Allowing them to directly perform file read / write, model loading, or weight injection may alter the model state due to missing paths, mismatched target aliases, incorrect call order, invalid parameter fields, or duplicate calls. Simply having the language model provide the model name does not confirm whether the name is registered, whether the corresponding file exists, or whether the configuration matches the backbone. Therefore, language understanding is suitable for semantic parsing and candidate action selection, while operations involving file status, registration item matching, and parameter loading still need to be constrained by the set of allowed actions and deterministic rules.

[0010] In summary, existing methods such as attribute analysis, supervised segmentation, general cue segmentation, lightweight parameter adaptation, language model parsing, and cross-frame memory propagation can address some issues in seismic target enhancement, mask prediction, parameter compression, instruction understanding, or cross-slice expansion. However, these techniques still suffer from insufficient integration: the trained target-specific adaptation parameter set lacks a unified binding with the normative categories, aliases, and model configurations; the natural language parsing results lack path, category, structure, and state verification with the actual weight loading; after loading the target-specific adaptation parameter set, the 2D cue segmentation and 3D propagation processes lack consistent task semantic constraints; and manual error correction on error slices may remain at the level of single-frame replacement, failing to establish conditional memory updates and bidirectional repropagation. Therefore, a technical solution is needed that can sequentially connect target-level lightweight adaptation, model registration, constrained agent routing, deterministic execution, and 3D memory error correction, clarifying the input, output, and anomaly handling relationships of each stage and reducing parameter mismatch and propagation errors in multi-target interactive interpretation. Summary of the Invention

[0011] The technical problem this invention aims to solve is: to establish independently loadable lightweight adaptation parameters for multiple types of seismic geological targets with significant morphological differences without updating the original weights of the pre-trained image encoder, thereby reducing inter-target interference during joint training; to ensure that natural language requests can be reliably mapped to a unique target category, model configuration, and adaptation weights; to prevent missing paths, unregistered targets, illegal structured outputs, inconsistent configurations, and duplicate injections from entering the segmentation process; and to extend the two-dimensional interactive results to three-dimensional slice sequences, correcting the overall results through memory write-back and bidirectional repropagation after local propagation deviations occur. This application provides a two-dimensional and three-dimensional seismic geological feature interactive recognition method based on task separation lightweight adaptation and agent routing to solve the above problems.

[0012] The technical solution adopted in this invention is: a two-dimensional and three-dimensional seismic geological feature interactive recognition method based on task separation, lightweight adaptation, and agent routing, which includes the following steps:

[0013] Step 1: Obtain labeled 2D seismic slices and 3D seismic volumes, construct a video target segmentation sequence from the 2D seismic slices and 3D seismic volumes, and perform differential preprocessing:

[0014] Acquire labeled 2D seismic slices and 3D seismic volumes; organize the 2D seismic slices into a single-frame sequence and unfold the 3D seismic volumes along a selected direction into an ordered sequence that maintains the slice order and numbering consistency; perform robust truncation, normalization, channel conversion, and size adjustment on the image amplitude, and apply bilinear interpolation to the image and nearest neighbor interpolation to the labeled mask to form data that can be read by the cue-based video target segmentation model;

[0015] Step 2: Establish independent training tasks for each of the seven types of geological targets and configure LoRA low-rank adaptation branches:

[0016] Using pre-trained SAM2 as a shared base, independent training tasks were established for seven types of geological targets: faults, stratigraphy, river channels, seismic facies, karst caves, chimneys, and salt bodies. LoRA low-rank adaptation branches were set in the selected attention projection layer of the image encoder, so that each type of target has a lightweight parameter update branch that can be trained independently, and the adaptation parameters of different targets are not merged.

[0017] Step 3: Train and save the target-specific adaptation parameter set:

[0018] Freeze the original parameters of the pre-trained image encoder, train the LoRA matrix, feature pyramid, memory attention, memory encoder, cue encoder and mask decoder corresponding to each target; construct training interactions using points, boxes and mask cues, and jointly optimize through mask prediction, region overlap, mask quality and target existence loss; after training, save the training parameters of each type of target as target-specific adaptation parameter sets that can be loaded independently.

[0019] Step 4: Establish the model registry and constrained agent routing:

[0020] For each target-specific adaptation parameter set, a registration item is established, including the canonical category, alias, model configuration, weight path, backbone structure, input resolution, and availability status. A constrained agent is composed of a local language model and a deterministic executor. The local language model converts natural language requests into restricted structured actions, and the deterministic executor sequentially verifies the data path, category mapping, registration status, configuration consistency, and duplicate loading status. Only after all verifications pass, the model configuration and target-specific adaptation parameter set uniquely corresponding to the target are loaded. For missing paths, unregistered targets, illegal structured outputs, inconsistent configurations, and duplicate loading, rejection, return, or reuse processing is performed respectively.

[0021] Step 5: Perform 2D interactive segmentation, 3D bidirectional propagation, and memory write-back error correction.

[0022] For 2D slices, receive positive points, negative points, bounding boxes, or coarse mask cues and output the target mask; for 3D seismic volumes, write the cues and masks on key slices into conditional memory and propagate them in both directions along the slice index; when the propagation results show omissions, over-segments, or drifts, supplementary cues are added to the error slices, a correction mask is generated, the correction mask is re-encoded into conditional memory, and then propagated again from the error slices to both sides until the continuity or boundary requirements set by the interpreter are met.

[0023] After the above processing, a two-dimensional target mask corresponding to the selected geological target can be obtained for two-dimensional slices; for three-dimensional seismic bodies, a continuous three-dimensional target mask sequence formed by bidirectional propagation and memory write-back error correction can be obtained.

[0024] The specific method for robustly truncating and normalizing the image amplitude in step (1) above is as follows: Robustly truncate abnormal amplitudes using the following formula:

[0025] (1)

[0026] In equation (1), The original amplitude, The amplitude after truncation. and The 0.5 and 99.5 percentiles of the effective amplitude of a single seismic image are respectively used as the lower and upper thresholds for amplitude robust truncation; then, linear normalization is performed and the image is quantized to eight-bit grayscale using the following formula:

[0027] (2)

[0028] In equation (2), For normalized pixel values, To prevent extremely small constants with a denominator of zero, the resulting single-channel image is copied into three identical channels; bilinear interpolation is used for the image, nearest-neighbor interpolation is used for the annotation mask, and all are uniformly adjusted to... Pixels; the same truncation, normalization, channel duplication, and size transformation rules are used in both the training and inference phases.

[0029] The LoRA low-rank adaptation branches set in step (2) above form the LoRA matrix adaptation positions; for any adapted linear layer, its output is:

[0030] (3)

[0031] In equation (3), For the input feature vector, This is the output feature vector of the linear layer. For frozen pre-trained weights, and For trainable low-rank matrices, Let be the rank of the low-rank matrix. This is the scaling factor.

[0032] The joint optimization objective in step (3) above is defined as:

[0033] (4)

[0034] In equation (4), For focus mask loss and focus parameters , The Dice loss and smoothing constant , For mask quality prediction and actual The regression loss between them There is a loss to achieve the goal.

[0035] In step (4) above, seven standard categories are established for the seven types of geological targets; each standard category Each registration item corresponds to a set of aliases. Model configuration Adaptive weight path Backbone structure Input resolution and availability status The mapping relationship is as follows:

[0036] (5)

[0037] In equation (5), Indicates access via the model registry Query No. Registration items obtained from each standardized category;

[0038] The local language model extracts only the data path, target representation, and required actions from the natural language request; it does not directly perform file read / write operations or weight injection. Its output for each round is limited to action identifiers. Parameter object and completion mark :

[0039] (6)

[0040] Deterministic actuator parsing structured output And based on the registry Sum check function Update execution status:

[0041] (7);

[0042] In equation (7), This indicates the next execution state obtained after the deterministic executor processes the structured output of this round. Fixed state transition function; verification function Check the fields and types, the existence and readability of data paths, the unique mapping of aliases to canonical categories, the availability status of registered items, the consistency of the backbone structure and input resolution, and whether the same adaptation parameter has been loaded.

[0043] In step (5) above: key slice image features extracted by the image encoder and feature pyramid The prompt code generated by the prompt encoder And current conditioned memory processed through memory attention Common input mask decoder Obtain key slices Target mask on:

[0044] (8)

[0045] In equation (8), This is an image feature extraction module consisting of an image encoder and a feature pyramid. As a key slice, Results obtained by applying the key slice Key slice image features; The prompt code is obtained after the encoder encodes the interactive prompt; Memorize based on current conditions; For memory and attention; For mask decoders; 2D tasks directly output ; 3D mission will And key slice image feature encoding as key slice conditional memory And propagate along the slice index increasing and decreasing directions respectively:

[0046] (9)

[0047] In equation (9), The total number of slice sequences, For slice index, For key slice index, and These represent the propagation operators in the two index directions, For the first Predictive masks for each slice; after checking the initial 3D results, error slices with omissions, over-scaling, or drift are identified. Add positive and negative prompts to obtain the correction mask. The mask is re-encoded as a conditional memory and propagated to both sides:

[0048] (10)

[0049] In equation (10), Index the selected error slice. This is an image feature extraction module consisting of an image encoder and a feature pyramid. Input For the first Error slice image, For the first The mask corrected by interactive prompts on each error slice For memory encoder, To be and The new condition memory is written back after encoding. and The meaning is the same as in formula (9). This indicates the merging of the propagation results in both the forward and reverse directions. To extract from slices The updated mask results are generated after repropagation in both directions.

[0050] For ease of explanation, the pre-trained segmentation base described in this invention preferably employs the SegmentAnything Model 2 (SAM2). SAM2 is a cue-based segmentation model that can receive point, box, or mask cues and utilizes a memory mechanism to propagate the target mask between ordered image frames. The lightweight adaptation described in this invention preferably employs Low-Rank Adaptation (LoRA). LoRA is a method that adjusts the model's feature representation by adding two trainable low-rank matrices to selected linear layers while keeping the original weights of the pre-trained linear layers frozen, thus requiring fewer additional parameters.

[0051] This invention does not take any single item from SAM2, LoRA, language model, or memory propagation as the inventive point, but establishes a complete collaborative technology chain that includes independent training based on geological targets, binding based on registration items, routing based on constrained agents, loading based on deterministic rules, and feeding back manual error correction to 3D propagation.

[0052] The beneficial effects of this application are:

[0053] 1. This invention uses a target-specific adaptation parameter set so that geological targets with significant morphological differences no longer share the same low-rank update set, reducing inter-target interference under weak cueing conditions, while keeping the pre-trained image encoder unchanged, which facilitates incremental expansion and independent maintenance based on target.

[0054] 2. This invention uses continuous constraints through a registry, constrained agent routing, and deterministic verification to convert aliases and paths in natural language into unique loadable model items; missing, unregistered, illegal, or duplicate calls are processed before weight injection, reducing the risk of human selection errors and generative outputs directly controlling the model state.

[0055] 3. This invention uses key slice prompts, memory propagation, error slice correction, and bidirectional repropagation to allow interpreters to repair 3D continuity with minimal interaction, avoiding slice-by-slice drawing; the error correction process and model version can be recorded for easy review. Attached Figure Description

[0056] Figure 1 This is a flowchart of the present invention.

[0057] Figure 2 This is an example diagram of loop-correction used for channel segmentation by 3D propagation and interpretation personnel; in the diagram: (a) setting initial positive point hints on representative seismic slices; (b) generating initial 2D channel segmentation results; (c) supplementing positive point hints to recover under-segmented areas; (d) obtaining corrected 2D segmentation results; (e) generating initial 3D channel segmentation results through cross-slice propagation; (f) selecting slices with poor segmentation results and adding negative point hints to remove over-segmented areas; (g) obtaining corrected 2D segmentation results; (h) re-propagating the corrected mask to generate the final refined 3D channel segmentation results.

[0058] Figure 3 The diagram shows a 3D multi-fault cross-slice propagation segmentation map with prompts and guidance. In the diagram: (a) multi-target point prompts are set on representative seismic slices; (b) 2D fault segmentation results; (c) 3D fault planes generated through cross-slice propagation.

[0059] Figure 4 The diagram shows a three-dimensional multi-layer cross-slice propagation segmentation map with prompts and guidance. In the diagram: (a) multiple target point prompts are set on a representative seismic profile; (b) two-dimensional layer segmentation results; (c) three-dimensional layer planes generated through cross-slice propagation.

[0060] Figure 5 The diagram shows a 3D multi-channel cross-slice propagation segmentation map with prompts and guidance; in the diagram: (a) multi-target point prompts are set on representative seismic slices; (b) a 2D channel mask; and (c) a 3D channel body generated through cross-slice propagation.

[0061] Figure 6 The diagram shows a 3D seismic phase propagation segmentation map with hints and guidance; in the diagram: (a) mask hints and point hints are set on representative seismic slices; (b) 2D seismic phase mask; (c) 3D seismic phase volume generated by propagation across slices.

[0062] Figure 7 The diagram shows a three-dimensional multi-cavity cross-slice propagation segmentation map with prompts and guidance. In the diagram: (a) on a representative seismic slice, prompts are set around multiple cavity targets; (b) a two-dimensional cavity mask; and (c) a three-dimensional cavity body generated through cross-slice propagation.

[0063] Figure 8 The diagram shows a 3D salt body propagating across slices with a hint-guided approach. In the diagram: (a) a hint is set inside the target salt body; (b) a 2D salt body mask; and (c) a 3D salt body generated through cross-slice propagation. Detailed Implementation

[0064] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the technical structure and executable steps of the present invention, and are not intended to limit the scope of protection to specific software names, data formats, or parameter values. While maintaining the correspondence between the target-specific adaptation parameter set—model registry—constrained agent routing—memory write-back—bidirectional repropagation, the specific pre-training base, local language model, and file format can be equivalently replaced. The specific process follows... Figure 1 As shown: First, a small number of training data are constructed according to the geological targets, and task-separated LoRA lightweight adaptation training is performed. The model configuration and the weight path of the target-specific adaptation parameter set are written into the model registry. Then, the local language model in the constrained agent parses the data path and target category in the natural language, and the deterministic executor matches and loads the unique registration item. Finally, point, box, or mask prompts are received to directly identify two-dimensional seismic slices, or to propagate three-dimensional seismic volumes from key slices to adjacent slices, forming two-dimensional or three-dimensional recognition results of the corresponding geological targets.

[0065] Example:

[0066] A method for interactive recognition of two-dimensional and three-dimensional seismic geological features based on task separation, lightweight adaptation, and agent routing includes the following steps:

[0067] Step 1: Construct a video target segmentation sequence from 2D seismic slices and 3D seismic volumes and perform differential preprocessing:

[0068] Acquire labeled 2D seismic slices or 3D seismic volumes; organize the 2D seismic slices into a single-frame sequence, and unfold the 3D seismic volumes along a selected direction into an ordered sequence that maintains the slice order and instance numbering; perform robust truncation, normalization, channel conversion, and size adjustment on the image amplitude, and apply bilinear interpolation to the image and nearest neighbor interpolation to the labeled mask to form data that can be read by the cue-based video target segmentation model.

[0069] This embodiment constructs seven types of geological target data, totaling 93,850 image-mask pairs, and organizes them into 294 video target segmentation sequences. The data composition of each target is shown in Table 1 below.

[0070] Table 1. Composition of Fine-tuning Data

[0071]

[0072] The 3D data is sliced ​​according to a predetermined direction, and the numbering of the same target instance remains stable within the sequence; faults, horizons, channels, seismic facies, and karst caves use 3D data sequences, while salt bodies and chimneys use 2D slices to form single-frame sequences. For any pixel location... The 0.5 percentile of the effective amplitude of a single image. and the 99.5th percentile And robustly truncate the abnormal amplitude using the following formula:

[0073] (1)

[0074] In equation (1), The original amplitude, The amplitude after truncation. and These are the 0.5 percentile and 99.5 percentile of the effective amplitude of a single seismic image, respectively. They are then linearly normalized and quantized to eight-bit grayscale using the following formula:

[0075] (2)

[0076] In equation (2), For normalized pixel values, To prevent extremely small constants with a denominator of zero, the resulting single-channel image is copied into three identical channels; bilinear interpolation is used for the images, nearest-neighbor interpolation is used for the annotation mask, and all are uniformly adjusted to... Pixels. The same truncation, normalization, channel duplication, and size transformation rules are used in both the training and inference phases.

[0077] Step 2: Establish independent training tasks for each of the seven types of geological targets and configure LoRA low-rank adaptation branches:

[0078] Using the pre-trained SAM2 (Segment Anything Model 2) as a shared base, independent training tasks are established for faults, stratigraphic horizons, river channels, seismic facies, karst caves, chimneys, and salt bodies. LoRA low-rank adaptation branches are set in the selected attention projection layer of the image encoder, so that each type of target has a lightweight parameter update branch that can be trained independently, and the adaptation parameters of different targets are not merged.

[0079] This embodiment uses a pre-trained SAM2 as the shared segmentation base, where the image encoder employs a Hiera-B+ with 80.8 million parameters. LoRA low-rank adaptation branches are set in the query-key-value joint projection and output projection of the 24 attention blocks, forming a total of 48 LoRA (Low-Rank Adaptation) adaptation positions. For any adapted linear layer, its output is:

[0080] (3)

[0081] In equation (3), For the input feature vector, This is the output feature vector of the linear layer. For frozen pre-trained weights, and For trainable low-rank matrices, Let be the rank of the low-rank matrix. This is the scaling factor. In this embodiment, we take... , , ,thus ;matrix Using Kaiming uniform initialization, the matrix Initialize to zero, so that the residuals of the LoRA low-rank adaptation branches are updated to zero at the start of training, without changing the pedestal output. Patch embedding, position embedding, multilayer perceptron of each attention block, and normalization parameters remain frozen.

[0082] Step 3: Train and save the target-specific adaptation parameter set:

[0083] After completing the independent training tasks and LoRA adaptation location configurations described above, seven training tasks were launched, each using only the data of the corresponding geological target. During training, the original parameters of the Hiera-B+ image encoder were frozen, and the LoRA matrix, feature pyramid, memory attention, memory encoder, cue encoder, and mask decoder for 48 adaptation locations were jointly trained. In each iteration, 15 frames were randomly selected from a sequence, with a maximum of 4 target instances per frame; all frames in the same sequence underwent consistent random horizontal flipping. Point cues and box cues were enabled with a probability of 0.5, and a ground truth mask was attached as a mask cue with a probability of 0.1.

[0084] The LoRA matrix and the other trainable modules use a grouped learning rate. The LoRA matrix uses... The initial learning rate was reduced to a value after cosine annealing. Feature pyramid and downstream module usage The initial learning rate and reduced to The optimizer is AdamW, and the weight decay for both groups is 0.1. The joint optimization objective is defined as:

[0085] (4)

[0086] In equation (4), For focus mask loss and focus parameters , The Dice loss and smoothing constant , For mask quality prediction and actual The regression loss between them The target has a loss. After training, the LoRA matrix corresponding to each target and the parameters of the trainable downstream modules are saved separately, forming seven sets of target-specific adaptation parameters that are not mixed. Each parameter set includes a matrix of 48 LoRA adaptation positions and parameters of trainable downstream modules such as feature pyramid, memory attention, memory encoder, cue encoder and mask decoder. The correspondence is shown in Table 2.

[0087] Table 2. Set of Adaptive Parameters for Seven Types of Geological Targets

[0088]

[0089] Step 4: Establish the model registry and constrained agent routing:

[0090] For each target-specific adaptation parameter set, a registration item is established, including the canonical category, alias, model configuration, weight path, backbone structure, input resolution, and availability status. A constrained agent is composed of a local language model and a deterministic executor. The local language model converts natural language requests into restricted structured actions, and the deterministic executor sequentially verifies the data path, category mapping, registration status, configuration consistency, and duplicate loading status. Only after all verifications pass is the model configuration and target-specific adaptation parameter set uniquely corresponding to the target loaded. For missing paths, unregistered targets, illegal structured outputs, inconsistent configurations, and duplicate loading, rejection, return, or reuse processing is performed respectively.

[0091] Seven normative categories were established for the seven target categories. Each normative category... Each registration item corresponds to a set of aliases. Model configuration Adaptive weight path Backbone structure Input resolution and availability status The mapping relationship is as follows:

[0092] (5)

[0093] In equation (5), the left side of the equal sign indicates that the registry is accessed through the model. Query No. The registration items are obtained from the specification categories. Taking the fault registration item as an example, the specification category is denoted as fault, the alias set includes "fault", "fracture", and "fault", and the model configuration points to the configuration using the Hiera-B+ image encoder. The weight path points to the fault-specific adaptation parameter file, and its availability is marked as "available". The other six categories use the same field structure and are each bound to their own weights; the same alias can only map to one canonical category.

[0094] The local language model extracts only the data path, target representation, and required actions from the natural language request; it does not directly perform file read / write operations or weight injection. Its output for each round is limited to an action identifier. Parameter object and completion mark :

[0095] (6)

[0096] Action identifiers are taken from a preset allowed set, including data loading, target category matching, model configuration reading, adaptation parameter loading, and missing information reporting. The deterministic actuator parses the structured output. and based on the registry Sum check function Update execution status:

[0097] (7);

[0098] In equation (7), This indicates the next execution state obtained after the deterministic executor processes the structured output of this round. This is a fixed-state transition function. Verification function. The system sequentially checks the fields and types, the existence and readability of data paths, the unique mapping from aliases to canonical categories, the availability status of registered items, the consistency of the backbone structure and input resolution, and whether the same adaptation parameters have been loaded. Valid requests are executed in the order of loading, matching, reading, and injection; requests with missing paths are rejected before weight loading; unregistered targets return "New training task required"; illegal structured outputs are discarded and the system enters the deterministic missing information branch; injection stops when configurations and weights are inconsistent; existing model instances are reused when the same category and parameter version have already been loaded.

[0099] Step 5: Perform 2D interactive segmentation, 3D bidirectional propagation, and memory write-back error correction.

[0100] After loading the target-specific adaptation parameter set, interpreters set positive points, negative points, bounding boxes, or coarse masks on representative slices of 2D or 3D seismic bodies to output the target mask. For 3D seismic bodies, the cues and masks on key slices are written into conditional memory and propagated in both directions along the slice index. When the propagation results show omissions, over-segments, or drift, cues are added to the error slices, a correction mask is generated, and this correction mask is re-encoded into conditional memory. Cross-slice propagation is performed along the slice sequence. Key slice image features are extracted by the image encoder and feature pyramid. The prompt code generated by the prompt encoder And current conditioned memory processed through memory attention Common input mask decoder Obtain key slices Target mask on: (8)

[0101] In equation (8), This is an image feature extraction module consisting of an image encoder and a feature pyramid. As a key slice, Results obtained by applying the key slice Key slice image features; The prompt code is obtained after the encoder encodes the interactive prompt; Memorize based on current conditions; For memory and attention; This is a mask decoder. The 2D task outputs directly. ; 3D mission will And key slice image feature encoding as key slice conditional memory And propagate along the slice index increasing and decreasing directions respectively:

[0102] (9)

[0103] In equation (9), The total number of slice sequences, For slice index, For key slice index, and These represent the propagation operators in the two index directions, For the first Predictive masks for each slice. After checking the initial 3D results, errors in slices that are under-segmented, over-segmented, or drifted are identified. Add positive and negative prompts to obtain the correction mask. The mask is re-encoded as a conditional memory and propagated to both sides:

[0104] (10)

[0105] In equation (10), Index the selected error slice. This is an image feature extraction module consisting of an image encoder and a feature pyramid. Input For the first Error slice image, For the first The mask corrected by interactive prompts on each error slice For memory encoder, To be and The new condition memory is written back after encoding. and The meaning is the same as in formula (9). This indicates the merging of the propagation results in both the forward and reverse directions. To extract from slices The updated mask results are generated after repropagation in both directions. Each error correction records the slice index, hint coordinates, hint positive / negative attributes, masks before and after correction, adapted parameter versions, and repropagation range. If errors still exist after repropagation, new error slices are selected for repeated hints, memory write-back, and bidirectional propagation until the continuity or boundary requirements set by the interpreter are met. The results are as follows: Figure 2 As shown:

[0106] Figure 2 Taking river identification as an example, this paper illustrates the complete process from key slice interactive segmentation, cross-slice propagation to manual error correction and memory write-back. Figure 2 In (a), the interpreters first selected representative slices with clear river channel features from the 3D seismic body and set initial point prompts inside the suspected river channel to indicate the target to be identified; Figure 2 In (b), the model generates an initial two-dimensional river channel mask based on the prompt, but due to the local weakening of the river channel reflection characteristics or unclear boundaries, there are omissions in the initial results; Figure 2 In (c), the interpreter checks the segmentation error, adds positive points at the missing segments, and guides the model to include the missing areas into the target range; Figure 2 In (d), the model recalculates the mask based on the added prompts, extending the 2D segmentation results further along the river channel boundary to obtain a more complete key slice river channel mask. Subsequently, the corrected 2D mask and its corresponding slice image features are encoded as conditional memory and propagated across slices in both the forward and reverse directions of the slice index, resulting in... Figure 2 (e) shows the initial three-dimensional channel segmentation results. After the interpreter reviewed the initial three-dimensional results, Figure 2(f) Select error slices that have been over-segmented or have drifted boundaries, and set negative points in the mis-segmented areas to clearly exclude non-channel parts; Figure 2 In (g), the model uses negative point cues to re-segment the slice, removes the erroneous extension region, and forms a corrected two-dimensional channel mask. Finally, the corrected mask is re-encoded and written into conditional memory, and propagated to both sides using the error slice as the new error correction benchmark, so that the erroneous results in adjacent slices are updated synchronously, ultimately obtaining... Figure 2 (h) shows the fine segmentation result of the three-dimensional river channel. Thus, the positive and negative prompts from the interpreter not only correct the current two-dimensional slice, but also participate in subsequent cross-slice propagation through conditional memory, realizing the linkage update of the overall three-dimensional result by the local error correction in two-dimensionality.

[0107] Step 6: 3D recognition results on real data:

[0108] Figures 3 to 8 The three-dimensional identification process for faults, stratigraphic horizons, river channels, seismic facies, karst caves, and salt bodies is shown separately. The different colors in each figure are only used to distinguish different target instances or seismic facies types and do not represent amplitude or geological attribute values.

[0109] Figure 3 Display the 3D recognition results of multiple faults. Figure 3 In (a) on a representative seismic slice, multiple target points are set for multiple faults. The location of the target points is selected near the fault traces where the reflection phase axis is discontinuous or broken, and independent target numbers are assigned to different fault instances. Figure 3 (b) in the figure gives the results of two-dimensional fault identification on the prompt slice. Each fault mask extends along the corresponding discontinuity zone, and the instances are still separated when the fault spacing is small. The tilt and bending morphology of the fault traces in the section are also depicted. Figure 3 In step (c), each two-dimensional fault mask is written into the conditional memory and propagated to both sides along the slice sequence to form independent three-dimensional fault planes. The results maintain the continuous extension of faults between adjacent slices and the spatial separation relationship of different fault planes, and can be used to characterize the three-dimensional distribution of multiple faults.

[0110] Figure 4 Display the results of multi-level 3D recognition. Figure 4 In (a) on a representative seismic profile, multiple target point prompts are set along different reflection phase axes so that each prompt corresponds to a layer to be tracked. Figure 4 (b) in the figure gives the two-dimensional layer identification results on the prompt profile. Each layer mask extends along the corresponding phase axis and maintains the separation between instances when the distance between adjacent layers is close. Figure 4In (c), the two-dimensional stratigraphic mask is propagated as a conditional memory along the slice sequence to form multiple independent three-dimensional stratigraphic planes. The propagation results maintain the vertical order, lateral continuity, and undulation morphology of each stratigraphic plane as it changes with spatial location, and can be used to simultaneously track multiple sets of stratigraphic planes.

[0111] Figure 5 The results of 3D recognition of multiple river channels are displayed. Figure 5 In (a) of the representative time slice, point prompts are set inside multiple river targets respectively, so that the prompts are associated with each river instance. Figure 5 (b) in the figure gives a two-dimensional river channel mask, and the identification results extend along the bends and strip-shaped reflection anomalies, while keeping adjacent or locally close river channels as different instances; Figure 5 In step (c), the river channel masks are propagated along the slice direction to obtain the corresponding three-dimensional river volume. The resulting three-dimensional results preserve the positional migration, width variation, and tortuous extension characteristics of the river channels on different slices, and can express the spatial distribution relationship of multiple river volumes.

[0112] Figure 6 The results of three-dimensional identification of different seismic phases are displayed. Figure 6 (a) Based on the morphological differences of seismic facies, interactive prompts are set: coarse mask prompts are used for progradational and retrogradational facies with lateral continuity and relatively clear regional boundaries, and positive and negative point combination prompts are used for chaotic facies with disordered texture and irregular boundaries, so as to define the target area and exclude adjacent non-target reflections respectively. Figure 6 (b) in the figure gives a two-dimensional seismic phase mask on a representative profile, with different masks corresponding to different types of reflective structures; Figure 6 In step (c), each 2D mask is written into the conditional memory and propagated across slices to form mutually distinguishable 3D seismic phases. This result demonstrates that different cue types can be selected based on the target geometry, and continuous phases and phases with irregular boundaries can be processed in the same identification process.

[0113] Figure 7 The results of three-dimensional identification of multiple karst caves are displayed. Figure 7 In (a) on a representative seismic profile, box prompts are set around multiple local anomalies to limit the search range for each cave instance; Figure 7 (b) shows the two-dimensional cave mask formed within each prompt box. The mask covers local abnormal reflections and suppresses the background outside the box. Figure 7 (c) propagates the cave masks along the slice sequence to obtain multiple spatially discrete three-dimensional cave bodies. During the propagation process, each cave instance is kept independently numbered to avoid merging closely spaced local anomalies into the same target. This method is suitable for geological targets that are small in size, compact in shape, and discretely distributed.

[0114] Figure 8The results of three-dimensional recognition of multi-salt bodies are displayed. Figure 8 (a) Set up time prompts inside different target salt bodies to specify the salt body instances that need to be identified; Figure 8 (b) in the figure provides a two-dimensional salt body mask on the hint profile, which extends from the hint location to the salt body boundary and defines the target range based on the reflection difference between the salt body and the surrounding rock; Figure 8 (c) In this method, multiple three-dimensional salt bodies are formed through cross-slice propagation. The results preserve the variation of salt body boundaries with slice position and the spatial separation relationship between different salt bodies, demonstrating that this invention can drive the three-dimensional recognition of targets in large-scale continuous regions with a small number of internal cues. Figures 3 to 8 It can be seen that the same "hint-2D mask-conditional memory-cross-slice propagation" process can be adapted to slender interfaces, curved strips, regional phases, discrete small targets and large-scale continuous geological bodies.

[0115] The above embodiments demonstrate that the contribution of the present invention to the prior art is as follows:

[0116] This invention forms an executable closed loop through the training and storage of a target-specific adaptation parameter set, registration item binding, constrained agent routing, deterministic loading, 2D / 3D cue segmentation, and bidirectional repropagation of error correction memory. Without deviating from the correspondence of this closed loop, equivalent adjustments to the number of target categories, slicing direction, allowed action names, file formats, and human-computer interaction termination conditions can all be implemented using the technical concept of this invention. Specifically:

[0117] 1. Establish a geological target-level task separation mechanism: train LoRA matrices and related downstream modules for seven types of targets respectively to form mutually independent target-specific adaptation parameter sets; each parameter set is then bound to the standard category, alias, configuration, backbone, resolution and status as a registration item, so that the training products and online loading objects have a verifiable one-to-one correspondence.

[0118] 2. Establish a division of labor mechanism between the local language model and the deterministic executor within the constrained agent: The local language model only outputs structured actions within the allowed set, while path verification, category normalization, registration item matching, configuration consistency, weight injection, and duplicate loading judgment are all completed by the deterministic executor, thus isolating the generative output from the model's internal state from a technical process perspective.

[0119] 3. The registration and loading results are used for both 2D and 3D prompt segmentation: 2D slices are directly and interactively identified, and 3D seismic bodies are propagated bidirectionally using key slices as conditional frames, so that the geological semantics and prompt response of the same target-specific adaptation parameter set remain consistent across different data dimensions.

[0120] 4. Establish a closed loop of error correction mask-conditional memory-two-way repropagation: After the error slice is manually corrected, the correction mask is re-encoded as conditional memory and propagated to both sides from the slice, so that a local error correction can update multiple adjacent slices instead of just covering a single frame result.

[0121] This invention can be used for two-dimensional and three-dimensional geological interpretation in oil and gas exploration and development, geothermal exploration, carbon dioxide geological storage, underground engineering, and seismic data research, including fault tracing, stratigraphic picking, channel characterization, seismic facies zoning, cave identification, chimney detection, and salt body boundary extraction; it can also be integrated into an offline geophysical intelligent interpretation platform to provide natural language model selection, prompt segmentation, cross-slice propagation, and manual error correction in a data-secure environment.

Claims

1. A method for interactive recognition of two-dimensional and three-dimensional seismic geological features based on task separation, lightweight adaptation, and agent routing, characterized in that: The method includes the following steps: Step 1: Obtain labeled 2D seismic slices and 3D seismic volumes, construct a video target segmentation sequence from the 2D seismic slices and 3D seismic volumes, and perform differential preprocessing: Acquire labeled 2D seismic slices and 3D seismic volumes; organize the 2D seismic slices into a single-frame sequence and unfold the 3D seismic volumes along a selected direction into an ordered sequence that maintains the slice order and numbering consistency; perform robust truncation, normalization, channel conversion, and size adjustment on the image amplitude, and apply bilinear interpolation to the image and nearest neighbor interpolation to the labeled mask to form data that can be read by the cue-based video target segmentation model; Step 2: Establish independent training tasks for each of the seven types of geological targets and configure LoRA low-rank adaptation branches: Using pre-trained SAM2 as a shared base, independent training tasks were established for seven types of geological targets: faults, stratigraphy, river channels, seismic facies, karst caves, chimneys, and salt bodies. LoRA low-rank adaptation branches were set in the selected attention projection layer of the image encoder, so that each type of target has a lightweight parameter update branch that can be trained independently, and the adaptation parameters of different targets are not merged. Step 3: Train and save the target-specific adaptation parameter set: Freeze the original parameters of the pre-trained image encoder, train the LoRA matrix, feature pyramid, memory attention, memory encoder, cue encoder and mask decoder corresponding to each target; construct training interactions using points, boxes and mask cues, and jointly optimize through mask prediction, region overlap, mask quality and target existence loss; after training, save the training parameters of each type of target as target-specific adaptation parameter sets that can be loaded independently. Step 4: Establish the model registry and constrained agent routing: For each target-specific adaptation parameter set, a registration item is established, including the canonical category, alias, model configuration, weight path, backbone structure, input resolution, and availability status. A constrained agent is composed of a local language model and a deterministic executor. The local language model converts natural language requests into restricted structured actions, and the deterministic executor sequentially verifies the data path, category mapping, registration status, configuration consistency, and duplicate loading status. Only after all verifications pass, the model configuration and target-specific adaptation parameter set uniquely corresponding to the target are loaded. For missing paths, unregistered targets, illegal structured outputs, inconsistent configurations, and duplicate loading, rejection, return, or reuse processing is performed respectively. Step 5: Perform 2D interactive segmentation, 3D bidirectional propagation, and memory write-back error correction. For 2D slices, receive positive points, negative points, bounding boxes, or coarse mask cues and output the target mask; for 3D seismic volumes, write the cues and masks on key slices into conditional memory and propagate them in both directions along the slice index; when the propagation results show omissions, over-segments, or drifts, supplementary cues are added to the error slices, a correction mask is generated, the correction mask is re-encoded into conditional memory, and then propagated again from the error slices to both sides until the continuity or boundary requirements set by the interpreter are met. After the above processing, a two-dimensional target mask corresponding to the selected geological target can be obtained for two-dimensional slices; for three-dimensional seismic bodies, a continuous three-dimensional target mask sequence formed by bidirectional propagation and memory write-back error correction can be obtained.

2. The two-dimensional and three-dimensional seismic geological feature interactive recognition method based on task separation, lightweight adaptation, and agent routing according to claim 1, characterized in that: The specific method for robustly truncating and normalizing the image amplitude in step 1 is as follows: Robustly truncate abnormal amplitudes using the following formula: ；（1） In equation (1), The original amplitude, The amplitude after truncation. and The 0.5 and 99.5 percentiles of the effective amplitude of a single seismic image are respectively used as the lower and upper thresholds for amplitude robust truncation; then, linear normalization is performed and the image is quantized to eight-bit grayscale using the following formula: ；（2） In equation (2), For normalized pixel values, To prevent extremely small constants with a denominator of zero, the resulting single-channel image is copied into three identical channels; bilinear interpolation is used for the image, nearest-neighbor interpolation is used for the annotation mask, and all are uniformly adjusted to... Pixels; the same truncation, normalization, channel duplication, and size transformation rules are used in both the training and inference phases.

3. The two-dimensional and three-dimensional seismic geological feature interactive recognition method based on task separation, lightweight adaptation, and agent routing according to claim 1, is characterized in that: The low-rank LoRA adaptation branches set in step 2 form the LoRA matrix adaptation positions; for any adapted linear layer, its output is: ；（3） In equation (3), For the input feature vector, This is the output feature vector of the linear layer. For frozen pre-trained weights, and For trainable low-rank matrices, Let be the rank of the low-rank matrix. This is the scaling factor.

4. The two-dimensional and three-dimensional seismic geological feature interactive recognition method based on task separation, lightweight adaptation, and agent routing according to claim 1, characterized in that: The joint optimization objective in step 3 is defined as follows: ；（4） In equation (4), For focus mask loss and focus parameters , The Dice loss and smoothing constant are , For mask quality prediction and actual The regression loss between them There is a loss to achieve the goal.

5. The two-dimensional and three-dimensional seismic geological feature interactive recognition method based on task separation, lightweight adaptation, and agent routing according to claim 1, characterized in that: Step 4 establishes seven standard categories for the seven types of geological targets; each standard category Each registration item corresponds to a set of aliases. Model configuration Adaptive weight path Backbone structure Input resolution and availability status The mapping relationship is as follows: ；（5） In equation (5), Indicates access via the model registry Query No. Registration items obtained from each standardized category; The local language model extracts only the data path, target representation, and required actions from the natural language request; it does not directly perform file read / write operations or weight injection. Its output for each round is limited to action identifiers. Parameter object and completion mark : ；（6） Deterministic actuator parsing structured output and based on the registry Sum check function Update execution status: （7）； In equation (7), This indicates the next execution state obtained after the deterministic executor processes the structured output of this round. Fixed state transition function; verification function Check the fields and types, the existence and readability of data paths, the unique mapping of aliases to canonical categories, the availability status of registered items, the consistency of the backbone structure and input resolution, and whether the same adaptation parameter has been loaded.

6. The two-dimensional and three-dimensional seismic geological feature interactive recognition method based on task separation, lightweight adaptation, and agent routing according to claim 1, characterized in that: Step 5: Key slice image features extracted by the image encoder and feature pyramid The prompt code generated by the prompt encoder And current conditioned memory processed through memory attention Common input mask decoder Obtain key slices Target mask on: (8) In equation (8), This is an image feature extraction module consisting of an image encoder and a feature pyramid. As a key slice, Results obtained by applying the key slice Key slice image features; The prompt code is obtained after the encoder encodes the interactive prompt; Memorize based on current conditions; For memory and attention; For mask decoders; 2D tasks directly output ; 3D mission will And key slice image feature encoding as key slice conditional memory And propagate along the slice index increasing and decreasing directions respectively: ；（9） In equation (9), The total number of slice sequences, For slice index, For key slice index, and These represent the propagation operators in the two index directions, For the first Predictive mask for each slice; After checking the initial 3D results, error slices were selected based on whether there was omission, over-scaling, or drift. Add positive and negative prompts to obtain the correction mask. The mask is re-encoded as a conditional memory and propagated to both sides: ；（10） In equation (10), Index the selected error slice. This is an image feature extraction module consisting of an image encoder and a feature pyramid. Input For the first Error slice image, For the first The mask corrected by interactive prompts on each error slice For memory encoder, To be and The new condition memory is written back after encoding. and The meaning is the same as in formula (9). This indicates the merging of the propagation results in both the forward and reverse directions. To extract from slices The updated mask results are generated after repropagation in both directions.

Images