A deep learning-based core lithology image intelligent recognition system

Abstract

The present application relates to a kind of core lithology image intelligent identification system based on deep learning, its core includes image acquisition module, image pre-processing module, feature extraction and classification module and data integration and output module four modules, through integrating high-precision image acquisition, AI noise reduction, multi-feature fusion classification and automatic report generation technology, through image intelligent analysis, the fast, accurate classification of core lithology is realized, can significantly improve the efficiency and accuracy of geological exploration, suitable for the lithology automatic classification and intelligent identification of core sample obtained in the process of soil hole and groundwater monitoring well drilling of soil sampling, can provide efficient and accurate technical support for environmental hydrogeological exploration.

Description

Technical Field

[0001] This invention relates to the interdisciplinary application of artificial intelligence technology and environmental hydrogeology, and in particular to an intelligent identification system for core lithology images based on deep learning. Background Technology

[0002] Core lithology identification is a core technical link in environmental hydrogeological exploration. The accuracy of its identification results is directly related to the reliability of groundwater monitoring systems and the scientific nature of engineering geological exploration. Traditional core lithology identification methods mainly rely on the visual observation and experience judgment of geological technicians. This manual identification mode has the following significant limitations: (1) Low identification efficiency: Traditional methods require layer-by-layer observation, description and recording of lithological characteristics of core samples. This process not only consumes a lot of manpower and resources, but also reduces work efficiency further when encountering complex rock strata structures (such as interlayers, interbeds or transition zones), significantly extending the project cycle; (2) High subjective dependence: Since it relies entirely on the professional experience of technicians, different personnel may have different judgment standards for the same core, resulting in a lack of consistency and repeatability of classification results, making it difficult to achieve standardized operations; (3) Insufficient data recording integrity: Manual recording methods easily overlook key details, such as the gradual transition of mineral composition, texture features and special structural marks. The lack of this information may affect the accuracy of subsequent geological analysis.

[0003] With the rapid development of deep learning technology, computer vision has demonstrated outstanding performance in fields such as image classification and object detection. However, the application of existing technologies in geological core lithology identification still faces challenges:

[0004] (1) Complex noise interference problem: During the process of groundwater drilling and coring, the core images obtained are often subject to a combination of complex noise interference, including but not limited to local contamination caused by mud adhesion, random noise introduced by dust particles, contrast imbalance caused by uneven illumination, and block effect and color banding distortion caused by image compression. These mixed noises not only obscure the key geological features of the core (such as mineral texture, fracture distribution and grain boundaries), but also significantly reduce the accuracy of subsequent analysis.

[0005] (2) Bottleneck of multimodal feature extraction: Accurate lithology identification requires comprehensive analysis of multidimensional features such as color distribution, texture features, mineral grain size, and structural morphology. This is different from traditional machine learning methods and improves the generalization ability of the model to a certain extent.

[0006] (3) Lack of standardized output: Existing tools lack the ability to automatically generate standardized groundwater monitoring well construction record sheets.

[0007] (4) Blank in soil lithology identification: Current research mainly focuses on rock lithology identification, and no effective solution has yet been formed for automated image identification of soil types (such as clay, sand, etc.).

[0008] (5) Lack of professional datasets: Existing rock image datasets generally lack systematic design for groundwater monitoring well drilling (hydrogeological) scenarios, especially in terms of coverage and annotation professionalism of typical hydrogeological core lithology samples.

[0009] To address the aforementioned issues, Chinese patent application publication number CN117874469A proposes a deep learning-based intelligent soil lithology identification method. This method classifies soil samples by analyzing their physicochemical properties (such as pH value and density) and combining them with environmental parameters. It focuses on multi-parameter quantitative analysis rather than the specific application scenario of automated classification of water-bearing soil and rock lithology during groundwater monitoring well drilling.

[0010] In addition, Chinese patent application publication number CN119540625A proposes a geological rock lithology identification and classification method and device. It adopts a general target detection process of "image classification model + region generation model + bounding box regression optimization". It improves the classification accuracy of conventional rock images through multi-scale feature extraction. It is a general system applicable to fields such as mineral exploration, rather than a special application field for hydrogeological drilling core images.

[0011] Therefore, it is currently very necessary to develop a system that is more suitable for intelligent identification of rock core lithology images in terms of data source selection, noise processing, multimodal information fusion and system specialization, so as to better meet the actual needs of efficient, accurate and standardized identification of rock core lithology in hydrogeological exploration. Summary of the Invention

[0012] The purpose of this invention is to address the problems of low efficiency, strong subjectivity, and incomplete data recording in current core identification, which mainly relies on manual experience. This invention proposes a deep learning-based intelligent core lithology image identification system (i.e., the EHG-lyzer system). By integrating high-precision image acquisition, AI noise reduction, multi-feature fusion classification, and automated report generation technologies, this system achieves efficient, objective, and standardized identification of core lithology, providing an intelligent solution for environmental hydrogeology.

[0013] To achieve the above objectives, this invention provides a deep learning-based intelligent identification system for core lithology images, comprising:

[0014] The image acquisition module acquires global core images using a portable high-resolution core image acquisition device, and acquires local core images at intervals of 1.2-1.5m. Simultaneously, it collects basic project information, professional technical indicators, and hydrogeological data as metadata on-site. The global images are ≥5 million pixels, and the local close-up images are ≥20 million pixels.

[0015] The image preprocessing module includes AI denoising, multi-noise adaptive processing, and detail enhancement. The AI ​​denoising utilizes PyTorch based on an improved U-Net / ResNet architecture to automatically distinguish between real details and noise in the image, eliminating mud contamination, dust noise, uneven lighting, and image distortion interference, while preserving the core texture and grain boundaries. The multi-noise adaptive processing includes cascaded processing of Gaussian noise, salt-and-pepper noise, and mixed noise, automatically adjusting algorithm parameters. The detail enhancement includes intelligently sharpening mineral grain boundaries and cracks during the denoising process.

[0016] The feature extraction and classification module accesses a professional geographic database, integrates regional stratigraphic information, automatically associates the regional geological background of core samples with GPS coordinates, stitches metadata and image features together, performs multimodal data fusion, and inputs it into a deep learning model. The deep learning model uses a convolutional neural network built on the PyTorch / TensorFlow framework combined with a professional geographic database to perform feature extraction, classification, and output.

[0017] The data integration and output module includes standardized report generation and database management. The standardized report generation is based on Python scripts to automatically fill in the groundwater monitoring well construction record form, and at the same time, it automatically generates a schematic diagram of the monitoring well structure based on the integrated Matplotlib / Plotly plotting library. The groundwater monitoring well construction record form includes basic information, technical parameters, hydrogeological data and lithological classification results. The database management uses PostgreSQL+PostGIS to store core images, classification results and geographic data.

[0018] Preferably, the AI ​​noise reduction is based on a convolutional neural network built on the PyTorch / TensorFlow framework, supports multi-task classification, and introduces residual connections and attention mechanisms.

[0019] Preferably, in performing multi-noise adaptive processing, multi-noise type processing includes:

[0020] Uneven lighting correction: Automatically balances shadow and highlight areas, correcting local overexposure / underexposure caused by insufficient ambient light or reflections;

[0021] Mud contamination remediation: Identify and eliminate mud-adhered areas on the surface of water drill cores, restoring the original color and texture;

[0022] Dust and noise removal: Remove dust and noise from dry drill cores;

[0023] Reduce photo distortion: Eliminate blockiness and banding, and restore high-frequency details.

[0024] Preferably, the multi-noise adaptive processing step includes:

[0025] Step S11, Pre-train the noise classification model: Use a lightweight CNN, input the local variance, frequency domain energy, and texture complexity features of the image, and output noise type labels;

[0026] Step S12: Based on the noise classification results, different branches are automatically activated: when the noise is Gaussian noise, frequency domain Wiener filtering is performed for enhancement; when the noise is salt-and-pepper noise, nonlocal mean suppression is performed; when the noise is mixed noise, cascaded multi-stage processing is performed, first using nonlocal mean or morphological filtering to eliminate mud contamination, and then using frequency domain Wiener filtering to process dust.

[0027] Step S13, Training Data and Loss Function: Collect core images under different drilling environments, label noise areas, and use generative adversarial networks to simulate complex noise data. Add MMD loss, fine-tune the generator to reduce the distribution gap between synthesized and real noise, use MS-SSIM to prevent over-smoothing, and use L1 norm to constrain noise residuals and hybrid loss functions.

[0028] Preferably, in the feature extraction and classification module, the training steps of the deep learning model include:

[0029] Step S21, Data preparation and enhancement: Prepare 20,000 to 30,000 images with more than 20 million pixels after denoising and synchronously recorded metadata, and perform data enhancement operations such as small-scale rotation, scaling and / or cropping on the images;

[0030] Step S22, Design the model architecture: Design a deep learning model using the PyTorch / TensorFlow framework and connect it to a database containing regional geological information. Input the denoised image and metadata to obtain the lithology category probability. Then, stitch the metadata to the fully connected layer for multimodal fusion and use MMD loss to perform domain adaptation training on the core image.

[0031] Step S23, Combating overfitting: Using spatial stratified sampling, the dataset is divided into training and validation sets based on the drilling location. If the accuracy of the validation set does not significantly improve within several consecutive training cycles, training is terminated.

[0032] Step S24, Model Validation: Establish a local core feature database, store the defined classification criteria, and associate it with regional stratigraphic information. Use the Torchcam library of PyTorch to generate activation heatmaps of CNN classification results, locate the key areas of interest of the model, match the heatmap features with the standard features in the local feature library, and verify the rationality of the model.

[0033] Preferably, the steps for automating the filling of groundwater monitoring well construction records based on Python scripts include:

[0034] Step S31, Data Acquisition and Preprocessing: Extract key parameters from the classification results output by the deep learning model as the basic data source for well construction records, combined with auxiliary data collected on site, including geographic coordinates, time, and borehole depth;

[0035] Step S32, Standardized Template Design: Design a template for the groundwater monitoring well construction record form, including the basic information area, technical parameter area, hydrogeological data, and graphic area;

[0036] Step S33: Use Python scripts to automate document generation: automatically fill template fields using a document processing library, match technical parameters according to classification results, and automatically generate groundwater monitoring well construction record forms that conform to industry standards.

[0037] Preferably, the information in the basic information area includes: project name, weather conditions, well number, drilling unit, drilling rig model and parameters, drilling parameters, well casing parameters, and well construction time; the information in the technical parameters area includes: filter pipe parameters, filter material type, well cover type, and well bottom sealing method; the information in the hydrogeological data area includes: surface elevation, initial water level depth, stable water level depth, and well flushing method.

[0038] Based on the above technical solution, the advantages of the present invention are:

[0039] The intelligent image recognition system for rock core lithology of the present invention (EHG-lyzer system) is based on deep learning technology. It achieves rapid and accurate classification of rock core lithology through intelligent image analysis, which can significantly improve the efficiency and accuracy of geological exploration. It is applicable to the automated classification and intelligent recognition of lithology of rock core samples obtained during the drilling of soil sampling boreholes and groundwater monitoring wells, and can provide efficient and accurate technical support for environmental hydrogeological exploration.

[0040] This invention innovatively uses high-resolution core images as the core data source, combines advanced image processing technologies (such as AI noise reduction and multi-feature fusion) to extract visual features, and integrates auxiliary information such as geographic coordinates and regional strata to achieve automated classification of water-bearing soil and rock lithology during the drilling of groundwater monitoring wells.

[0041] This invention addresses the unique characteristics of hydrogeological drilling core images by constructing a specialized technical system: the core algorithm integrates an adaptive AI noise reduction module to effectively handle the unique problems of mud contamination and uneven lighting in water / dry drilling cores, and uses multimodal fusion technology to deeply fuse images with geographic coordinates, borehole depth, and regional geological data, significantly improving classification accuracy in complex environments. Attached Figure Description

[0042] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0043] Figure 1 This is a schematic diagram of an intelligent identification system for rock core lithology images;

[0044] Figure 2 A flowchart illustrating the steps of adaptive multi-noise processing;

[0045] Figure 3 A flowchart illustrating the training steps of a deep learning model;

[0046] Figure 4 A step-by-step diagram illustrating the steps for automating the filling of groundwater monitoring well construction records using Python scripts;

[0047] Figure 5 This is a standard form for the well construction record sheet for groundwater monitoring wells in the embodiments. Detailed Implementation

[0048] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0049] This invention provides a deep learning-based intelligent identification system for core lithology images. It features significant innovations in data source selection, noise processing, multimodal information fusion, and system specialization, better meeting the practical needs of efficient, accurate, and standardized identification of core lithology in hydrogeological exploration. Its core comprises four modules: image acquisition, image preprocessing, feature extraction and classification, and data integration and output. Figure 1 As shown, it specifically includes:

[0050] (1) Image acquisition module

[0051] 1) A portable high-resolution core image acquisition device (global image ≥ 5 million pixels, local close-up image ≥ 20 million pixels) acquires global core images and acquires local core images at intervals of 1.2-1.5m, referring to the length of the core tube during on-site drilling.

[0052] To ensure image quality meets recognition requirements, the resolution of the global core image needs to be at least 5 million pixels, while the resolution of the local close-up core image needs to be at least 20 million pixels. The acquired images can be stored in multiple formats such as RAW or TIFF to meet the needs of subsequent image processing, analysis, and long-term preservation.

[0053] The image acquisition module can use a portable core image acquisition instrument to carry out the acquisition work. This instrument has good portability and ease of operation, can adapt to different field acquisition scenarios, and the image can reach 40 million pixels.

[0054] In this invention, global core images display the overall structure and macroscopic features of the core. For example, during the drilling of groundwater monitoring wells, the core may contain multiple lithological segments. Global images can clearly show the distribution, thickness variations, and transitional relationships between these segments. Simultaneously, they provide location and reference background for local close-up images. Local close-up core images can capture the microscopic details of the core, such as the size, shape, and arrangement of mineral grains, as well as the microstructure and texture of the rock, providing high-resolution detail information. This enables deep learning models to extract features more accurately and perform lithological classification.

[0055] 2) Collect metadata such as geographic coordinates, time, and borehole depth on-site simultaneously.

[0056] (2) Image preprocessing module

[0057] PyTorch can directly implement core algorithms such as AI denoising and enhancement, but it needs to be combined with other libraries in the Python ecosystem (such as OpenCV, PIL, and NumPy) to assist in data loading, preprocessing, and visualization to achieve the following functions:

[0058] 1) AI noise reduction: Using PyTorch based on an improved U-Net / ResNet architecture, interference such as mud contamination, dust noise, and uneven lighting is eliminated, while preserving the texture and particle boundaries of the core.

[0059] Deep learning-driven: Automatically distinguishes real details from noise (such as mud spots and dust particles) in images through convolutional neural networks (CNN), accurately eliminating interference while preserving key features such as core texture and mineral boundaries.

[0060] Furthermore, the AI-powered intelligent noise reduction function is achieved through the following core technologies, combining deep learning and image processing algorithms to accurately separate noise from real details:

[0061] Convolutional Neural Network (CNN) Architecture Design

[0062] A convolutional neural network (CNN) consists of convolutional layers, pooling layers, and fully connected layers.

[0063] Convolutional layers: mainly used for local feature extraction.

[0064] Pooling layer: mainly used for dimensionality reduction, reducing computation and memory usage.

[0065] Fully connected layer: The extracted features are flattened and then input into a traditional neural network for classification or regression.

[0066] In PyTorch, the core network architecture can be implemented using modified U-Net or ResNet, and its design features include:

[0067] 1. Multi-scale feature extraction (U-Net)

[0068] The encoder (downsampling) captures the global noise distribution (such as large-area contamination of mud), while the decoder (upsampling) recovers local details (such as the edges of mineral particles).

[0069] 2. Residual Learning (ResNet)

[0070] Convolutional networks directly predict noisy images, rather than clean images themselves, thus improving training efficiency.

[0071] ;

[0072] According to residual networks, the network only needs to learn the distribution residuals of noise, rather than the entire clean image.

[0073] 2) Multi-noise adaptive processing: Supports cascaded processing of Gaussian noise (frequency domain Wiener filtering), salt-and-pepper noise (non-local mean filtering), and mixed noise. Preferably, during multi-noise adaptive processing, the processing of multiple noise types includes:

[0074] Uneven lighting correction: Automatically balances shadow and highlight areas, correcting local overexposure / underexposure caused by insufficient ambient light or reflections.

[0075] Mud contamination remediation: Identify and eliminate mud-adhered areas on the surface of water drill cores, restoring the original color and texture.

[0076] Dust particle removal: Remove dust and noise from dry drill cores to avoid the particle effect on subsequent analysis.

[0077] Reduce photo distortion: Eliminate block artifacts and color banding, and restore high-frequency details.

[0078] like Figure 2 As shown, the steps of the multi-noise adaptive processing include:

[0079] Step S11, Pre-train the noise classification model: Use a lightweight CNN, taking the local variance, frequency domain energy, and texture complexity features of the input image, and outputting a noise type label. For example:

[0080] Gaussian noise has a uniform energy distribution in the frequency domain and small local variance fluctuations.

[0081] Salt and pepper noise: Localized sudden increase in variance (extremely bright / extremely dark pixels).

[0082] Mud pollution: low-frequency energy is high, and the continuity of texture is disrupted.

[0083] Step S12: Based on the noise classification results, different branches are automatically activated: when the noise is Gaussian noise, frequency domain Wiener filtering is performed for enhancement; when the noise is salt-and-pepper noise, nonlocal mean suppression is performed; when the noise is mixed noise, cascaded multi-stage processing is performed, first using nonlocal mean or morphological filtering to eliminate mud contamination, and then using frequency domain Wiener filtering to process dust.

[0084] Furthermore, based on the noise classification results, different branches are automatically activated, specifically:

[0085] 1. Gaussian noise → enhanced by frequency domain Wiener filter.

[0086] A. Convert the image to the frequency domain to obtain its frequency domain representation. In the frequency domain, the spectrum is corrected according to the following formula:

[0087]

[0088] Degradation function, which is a mathematical model describing the quality degradation of an image or signal during acquisition, transmission, and storage (fuzzy kernel, or 1 if there is no fuzziness).

[0089] The power spectral density ratio of noise to signal.

[0090] B. Restore the denoised image.

[0091] 2. Salt and pepper noise → Nonlocal mean (NLM) suppression.

[0092] For each pixel, search for all similar regions in the image and then replace the current pixel with a weighted average of the center pixel values ​​of these similar regions. The weighting formula is as follows:

[0093]

[0094] p, q: pixel position; h: smoothing parameter, which controls the adjustment coefficient for the trade-off between fidelity and smoothness during denoising or restoration.

[0095] Salt-and-pepper noise points (extremely bright / extremely dark) will be covered by values ​​from similar neighborhoods.

[0096] 3. Mixed noise → cascaded multi-stage processing

[0097] First, nonlocal mean (NLM) or morphological filtering is used to eliminate mud contamination (low-frequency noise), and then frequency domain Wiener filtering is used to process dust (high-frequency noise). The filtering parameters can be dynamically adjusted according to the noise intensity: mud region: increase the search window of NLM; dust region: increase the frequency domain cutoff threshold of Wiener filtering.

[0098] Step S13, Training Data and Loss Function: Core images from different drilling environments are collected, noisy areas are labeled, and a generative adversarial network (GAN) is used to simulate complex noise data. MMD loss is added, and the generator is fine-tuned to reduce the distribution gap between synthesized and real noise. MS-SSIM is used to prevent over-smoothing, and L1 norm is used to constrain noise residuals. A hybrid loss function is applied. Details are as follows:

[0099] 1. Data Synthesis Mode

[0100] Real noise data: Collect core images under different drilling environments (water drilling / dry drilling) and mark noise areas (such as mud masks).

[0101] Synthetic noise data: Generative Adversarial Networks (GANs) are used to simulate complex noise data. MMD loss is incorporated, and the generator is fine-tuned to reduce the distribution gap between the synthesized and real noise. The formula is as follows:

[0102] ;

[0103] Adversarial loss in GANs; Balance weights.

[0104] 2. Mixed loss function

[0105] MS-SSIM (Multi-Scale Structural Similarity) is used to prevent over-smoothing, and the L1 norm (also known as Manhattan distance or the first norm, which is the sum of the absolute values ​​of a vector element) is used to constrain the noise residuals, as shown in the following formula:

[0106] ;

[0107] Dynamic adjustment , Balancing detail with noise reduction intensity.

[0108] 3) Detail enhancement technology: Key features such as cracks and particle edges are enhanced through sharpening algorithms to prevent blurring.

[0109] 4) One-click batch noise reduction: Process hundreds of core scan images simultaneously while maintaining parameter consistency.

[0110] After using PyTorch to denoise the core images, the saved formats include TIFF, PNG, JPEG2000, and DNG.

[0111] (3) Feature extraction and classification module

[0112] 1) Multimodal data fusion:

[0113] It accesses a professional geographic database, integrates regional stratigraphic information (such as stratigraphic age and sedimentary environment), and automatically associates the regional geological background of core samples with GPS coordinates.

[0114] Metadata (drill depth, coordinates) is concatenated with image features (color, texture, grain size) and then input into a deep learning model.

[0115] 2) Deep learning models:

[0116] A convolutional neural network (CNN) is built based on the PyTorch / TensorFlow framework to support multi-task classification (rock / soil). Residual connections and attention mechanisms are introduced to improve the classification accuracy of complex lithology.

[0117] Furthermore, the feature extraction and classification module of this invention utilizes a convolutional neural network built on the PyTorch / TensorFlow framework combined with a professional geographic database to achieve in-depth learning for feature extraction, classification, and output.

[0118] like Figure 3 As shown, in the feature extraction and classification module, the training steps of the deep learning model include:

[0119] Step S21, Data Preparation and Enhancement: Prepare 20,000 to 30,000 images with more than 20 million pixels after denoising, as well as synchronously recorded metadata (such as borehole depth and GPS coordinates), and perform data enhancement operations such as small-scale rotation, scaling and / or cropping on the images.

[0120] Performing small-scale data augmentation operations such as rotation, scaling, and cropping on images can increase the robustness of the model to different image transformations, and deep learning models should, in principle, prioritize the use of real-world scene data.

[0121] Step S22, Design the model architecture: Design a deep learning model using the PyTorch / TensorFlow framework and connect it to a database containing regional geological information. Input the denoised image and metadata to obtain the lithology category probability. Then, concatenate the metadata to the fully connected layer for multimodal fusion and use MMD loss to perform domain adaptation training on the core image to reduce data distribution differences.

[0122] Step S23, Combating overfitting: Using a spatial stratified sampling method, the dataset is divided into a training set and a validation set based on the borehole location. If the accuracy of the validation set does not significantly improve within several consecutive training cycles, training is terminated.

[0123] Specifically, to ensure a reliable evaluation of the model's generalization performance, this invention employs a spatial stratified sampling method, dividing the dataset into training and validation sets based on borehole locations. During training, the validation set accuracy is monitored in real time and serves as a key indicator of model performance. If the validation set accuracy does not significantly improve within several consecutive training epochs (e.g., using early stopping), training is terminated to prevent overfitting.

[0124] Step S24, Model Validation: Establish a local core feature database, store the defined classification criteria (such as color, texture, grain size, and dry / water drill core shape features), and associate it with regional stratigraphic information (associating spatial data through the PostGIS geographic database). Use the Torchcam library of PyTorch to generate activation heatmaps of CNN classification results, locate the key areas of interest of the model, match the heatmap features with the standard features in the local feature library, and verify the rationality of the model.

[0125] When the difference between the model classification result and the feature library standard exceeds a threshold (the threshold range is dynamically optimized based on historical data statistics, business needs, and human feedback, rather than a fixed value), the following process is automatically triggered: ① Record the image, depth, and coordinates of the conflicting sample; ② Push an alarm through the web interface or mobile terminal to prompt geological experts to review; ③ After manual review, update the feature library or retrain the model.

[0126] In the deep learning-based intelligent identification system for rock core lithology images, the specific operation steps of the feature extraction and classification module are as follows:

[0127] 1. Open the EHG-lyzer software, create a new project and import auxiliary data. Bind the core image with data such as borehole depth, GPS coordinates, and logging curves to construct multi-dimensional input features (unify different data into the same coordinate system through depth alignment).

[0128] 2. The EHG-lyzer software accesses a database containing regional geological information, which can be used to make preliminary judgments on regional stratigraphic information based on GPS coordinates.

[0129] 3. Select the preprocessed core image and perform geometric correction and color calibration on the denoised core scan image to ensure that the image input to the deep learning model is distortion-free and has uniform illumination.

[0130] 4. Using EHG-lyzer, preliminary classification of global core image scans is performed, and more accurate classification is achieved by combining regional stratigraphic information with key features of local core images.

[0131] This invention develops a deep learning-based convolutional neural network (CNN) model using Python scripts for lithology classification of rock core images. The model's training dataset contains 20,000-30,000 rigorously selected and pre-processed field core photographs. These photographs have undergone denoising to improve image quality and are accurately labeled according to the classification criteria of this invention. To ensure the model can learn the common structural features of various types of rock cores, the training dataset covers diverse samples under various geological environments, lighting conditions, and coring techniques (such as water drilling and dry drilling).

[0132] Furthermore, the technical solution of this invention also incorporates a database containing regional geological information, further enriching the model's input data. This comprehensive and balanced dataset design enables the model to effectively capture the inherent patterns of key features such as color, texture, grain size, and shape in core images, thereby significantly improving the accuracy of lithological classification and the model's generalization ability.

[0133] (4) Data integration and output module

[0134] 1) Standardized report generation:

[0135] This system uses Python scripts to automatically populate the "Groundwater Monitoring Well Construction Record Form," including basic information (well number, drilling rig parameters), technical parameters (filter type, well washing method), hydrogeological data (water level depth), and lithological classification results. It also automatically generates a schematic diagram of the monitoring well structure (using Matplotlib / Plotly plotting libraries).

[0136] like Figure 4 As shown, preferably, the steps for automatically filling the groundwater monitoring well construction record form based on Python scripts include:

[0137] Step S31, Data Acquisition and Preprocessing: Extract key parameters from the classification results (CSV / JSON format) output by the deep learning model as the basic data source for well drilling records, combined with auxiliary data collected on site, including geographic coordinates, time, and borehole depth.

[0138] Step S32, Standardized Template Design: Design a template for the groundwater monitoring well construction record sheet, including the basic information area, technical parameter area, hydrogeological data, and graphic area.

[0139] Basic Information Area: Includes key fields such as project name, weather conditions, well number (including coordinate information), drilling unit, drilling rig model and parameters (drilling tools, drilling methods), borehole parameters (depth, diameter), well casing parameters (diameter, total length, wellhead height), and well construction time.

[0140] Technical Parameters Section: Detailed records of filter pipe parameters (length, type, diameter), filter media type, well cover type, well bottom sealing method, and other professional technical indicators.

[0141] Hydrogeological data: monitoring data such as surface elevation, initial water level depth, stable water level depth, and well flushing methods.

[0142] Illustration area: Automatically embeds "Schematic diagram of monitoring well structure and parameters" using Python scripts to automatically fill in the above information and automatically generate standardized well construction record forms based on the obtained classification results.

[0143] Step S33: Use Python scripts to automate document generation: automatically fill template fields using a document processing library, match technical parameters according to classification results, and automatically generate groundwater monitoring well construction record forms that conform to industry standards.

[0144] Furthermore, the system of this invention also supports batch generation and version management of record forms. For example... Figure 5 As shown, a schematic diagram of a groundwater monitoring well construction record sheet is presented.

[0145] 2) Database Management:

[0146] It uses PostgreSQL+PostGIS to store core images, classification results and geographic data, and supports fast retrieval by depth / coordinate.

[0147] In this invention, the main categories of soil core features that can be identified using deep learning models are as follows:

[0148] 1) Color

[0149] Crushed stone soil: The core material is usually grayish-white, light yellow, or brown (depending on the parent rock type; for example, granite crushed stone is grayish-white, while sandstone crushed stone is yellowish-brown). The particles are coarse, with visible gravel (particle size > 2mm), and mixed with sand or clay. The color does not change significantly after wetting, but the clay-filled parts may darken slightly (e.g., brown deepens). It has extremely high permeability, with surface water penetrating quickly and no stickiness.

[0150] Sandy soil: The core sample is mainly light-colored (grayish-white, pale yellow, light brown), with quartz sand mostly white and iron-containing sand possibly rust-yellow. The particles are loose, non-cohesive, and make a rustling sound when rubbed by hand. The color darkens slightly when moistened (e.g., from light yellow to yellowish-brown), but remains relatively light overall. Surface moisture seeps away quickly, leaving no luster, and it cannot be formed into a ball.

[0151] Silt: The core material is grayish-yellow, grayish-brown, or light gray in color, with fine particles like flour. It easily generates dust when dry. When it contains organic matter, it may appear darker (e.g., grayish-black). When moistened, the color darkens significantly (grayish-yellow → dark brown, gray → dark gray), with a slight sheen on the surface. It can be briefly kneaded into a ball, but crumbles easily when lightly pressed.

[0152] Clay: The core sample is predominantly dark in color (reddish-brown, dark gray, blackish-brown), with iron-containing clay appearing red and organic clay nearly black. When dry, it hardens into lumps and may crack into polygonal fissures. When moistened, the color becomes significantly darker and more uniform (e.g., reddish-brown → dark red, dark gray → blackish-gray), and the surface becomes glossy and smooth. It has strong viscosity and high plasticity, and can be rolled into thin strips or pressed into thin sheets.

[0153] 2) Particle size and particle characteristics

[0154] Particle Name Particle size (mm) Particle characteristics Boulder (stone) >200 Huge rock Pebbles (gravel) 200～20 Rounded or angular, about the size of a fist Gravel (pebbles) 20～2 Soybean to peanut size sand 2～0.075 Visible granules, similar to granulated sugar Powder 0.075～0.002 Fine as flour, but prone to dust when dry. clay <0.002 ultrafine particles

[0155] 3) Core shape characteristics

[0156] name Dry drill core shape characteristics Water drill core shape characteristics gravel soil The rock core was fragmented. The core sample was in the form of broken pieces with a small amount of mud adhering to the surface. sand Rock cores are easily scattered and lack cohesion. Rock cores are difficult to shape, and surface sand particles easily slip off, lacking cohesiveness. silt The core sample is short and columnar but easily fractured. The core samples are short columnar or blocky, but easily fractured, and have a gelatinized layer on the surface. clay The rock core is long and columnar with high plasticity. The core samples are long columnar or pasty in shape, with a smooth, glossy surface and a significantly darker color.

[0157] Furthermore, the rock core features that can be identified using deep learning models in this invention are categorized as follows:

[0158] 1) Texture and Structure

[0159] The texture and structure of rocks are determined by their formation process, mineral composition, and geological processes. The following are some typical rock structures.

[0160] Control factors Effects on texture and structure Typical example sedimentary environment Water flow velocity and transport distance determine grain sorting and bedding development (e.g., river sandstone and conglomerate have cross-bedding, while lacustrine mudstone has horizontal bedding). Alluvial fan conglomerate (disorderly deposition) vs. beach sandstone (parallel bedding). Mineral composition High quartz / feldspar content → granular support structure; abundant clay minerals → matrix support structure. Quartz sandstone (dense grains) vs. argillaceous sandstone (muddy filling the pores). Adhesion type Siliceous cement → dense and massive; calcareous cement → porous; clay-based cement → brittle. Siliceous cemented sandstone (high strength) vs. argillaceous cemented conglomerate (loose). Weathering and fissures Weathering weakens the bonds between particles, creating secondary pores; structural fissures increase permeability channels. Weathered granite (with a network of fissures) vs. unweathered basalt (with columnar joints). Diagenesis The degree of compaction and cementation affects porosity and structural integrity (e.g., loose sandstone with weak diagenesis). Consolidated sandstone (low porosity) vs. semi-consolidated conglomerate (high porosity).

[0161] 2) Color and gloss

[0162] In unconfined aquifers, the color and luster of rocks are mainly determined by their mineral composition, structural characteristics, degree of weathering, and type of cement, with significant differences among different rock types.

[0163] Influencing factors Effects on color and gloss Typical example Mineral composition Different minerals exhibit specific colors, such as quartz (colorless), hematite (reddish-brown), and clay minerals (grayish-white). Granite (feldspar → flesh-colored, quartz → vitreous luster); Sandstone (hematite cementation → reddish-brown) cement type Siliceous cement (grayish-white, hard and lustrous), calcareous cement (white, dull), ferrous cement (reddish-brown, metallic luster). Siliceous conglomerate (glassy luster); ferruginous sandstone (dark red, greasy luster) degree of weathering Weathering alters minerals (e.g., feldspar → kaolinite), making them lighter in color (grayish-white) and changing their luster from vitreous to earthy. Strongly weathered granite (grayish-white, matte); unweathered basalt (black, vitreous luster). Structural features Grain size and arrangement affect reflected light: coarse-grained rocks (such as conglomerate) have uneven luster, while fine-grained rocks (such as mudstone) have a more uniform surface. Coarse sandstone (strong granular texture, dull); limestone (dense, waxy luster). Water content When moistened, the color deepens (especially in clay minerals), and the luster increases (due to water film reflection). Claystone (grayish-white and dull when dry, dark gray and glossy when moist)

[0164] 3) Core characteristics and structural integrity

[0165] The state of a rock core is determined by the degree of weathering, the properties of the rock itself, and environmental conditions.

[0166] ① Intensity of weathering

[0167] Core integrity: quantitatively characterized by RQD (Rock Quality Index). Highly weathered cores are fragmented, while weakly weathered cores mostly maintain columnar integrity.

[0168] ② Properties of the rock itself

[0169] Mineralogical composition: The content of ferromagnesian minerals determines the base color, such as pyroxene causing a dark green tone, and feldspar showing a light color.

[0170] Structural characteristics: Columnar structures are commonly found in volcanic rock facies such as basalt, while massive structures are commonly found in plutonic rock bodies.

[0171] ③ Environmental conditions

[0172] The chemical effects of groundwater: Water molecules disrupt the structure of minerals (for example, feldspar turns into clay), causing rocks to soften or disintegrate; certain ions in groundwater (such as Na⁺ and Ca²⁺) replace the original ions in minerals, changing the strength or expansibility of rocks.

[0173] It should be noted that the core lithology image intelligent recognition system (EHG-lyzer system) of this invention is a scenario-specific system, specifically designed for lithology recognition of core samples from groundwater monitoring wells in the field of hydrogeology, unlike general-purpose systems used in conventional geological exploration scenarios (such as mineral exploration and geological hazard assessment). It is designed specifically for the unique characteristics of core images (noise, multimodal data) and automates the entire process from recognition to standardized output, generating standardized well drilling records by combining classification results and auxiliary information.

[0174] The EHG-lyzer system of this invention supports formats such as TIFF and PNG. The unique advantages of the EHG-lyzer system based on this invention in specialized geological data processing and multimodal data fusion are explained below:

[0175] (1) The EHG-lyzer system is designed for drilling core images of groundwater monitoring wells for hydrogeology professionals. It supports high-resolution scanning formats (such as TIFF / RAW) and can directly read metadata such as borehole depth and geographic coordinates, avoiding information loss during data conversion.

[0176] (2) Image data (color, texture) and non-image data (drilling depth, GPS coordinates, regional stratigraphic information) are concatenated into a unified input vector to realize the fusion of image and non-image data for classification by deep learning models, and support SQL queries. Combined with PostGIS spatial functions, geofence can be quickly retrieved.

[0177] (3) Each functional module (image preprocessing, classification model, report generation) exposes an interface through Python API, supports calls from other scripts or toolchains, and seamlessly connects with geological modeling software (such as GOCAD) and hydrological analysis tools (such as MODFLOW) to achieve full-link automation from core identification to simulation analysis.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A deep learning-based intelligent identification system for core lithology images, characterized in that: include: The image acquisition module acquires global core images using a portable high-resolution core image acquisition device, and acquires local core images at intervals of 1.2-1.5m. Simultaneously, it collects basic project information, professional technical indicators, and hydrogeological data as metadata on-site. The global image must be at least 5 million pixels, and the close-up image must be at least 20 million pixels. The image preprocessing module includes AI denoising, multi-noise adaptive processing, and detail enhancement. The AI ​​denoising utilizes PyTorch based on an improved U-Net / ResNet architecture to automatically distinguish between real details and noise in the image, eliminating mud contamination, dust noise, uneven lighting, and image distortion interference, while preserving the core texture and grain boundaries. The multi-noise adaptive processing includes cascaded processing of Gaussian noise, salt-and-pepper noise, and mixed noise, automatically adjusting algorithm parameters. The detail enhancement includes intelligently sharpening mineral grain boundaries and cracks during the denoising process. The feature extraction and classification module accesses a professional geographic database, integrates regional stratigraphic information, automatically associates the regional geological background of core samples with GPS coordinates, stitches metadata and image features together, performs multimodal data fusion, and inputs it into a deep learning model. The deep learning model uses a convolutional neural network built on the PyTorch / TensorFlow framework combined with a professional geographic database to perform feature extraction, classification, and output. The data integration and output module includes standardized report generation and database management. The standardized report generation is based on Python scripts to automatically fill in the groundwater monitoring well construction record form, and at the same time, it automatically generates a schematic diagram of the monitoring well structure based on the integrated Matplotlib / Plotly plotting library. The groundwater monitoring well construction record form includes basic information, technical parameters, hydrogeological data, and lithological classification results. The database management uses PostgreSQL and PostGIS to store core images, classification results, and geographic data. Among them, the processing of multiple noise types during multi-noise adaptive processing includes: Uneven lighting correction: Automatically balances shadow and highlight areas, correcting local overexposure / underexposure caused by insufficient ambient light or reflections; Mud contamination remediation: Identify and eliminate mud-adhered areas on the surface of water drill cores, restoring the original color and texture; Dust and noise removal: Remove dust and noise from dry drill cores; Reduce photo distortion: Eliminate blockiness and banding, and restore high-frequency details; The steps of the multi-noise adaptive processing include: Step S11, Pre-train the noise classification model: Use a lightweight CNN, input the local variance, frequency domain energy, and texture complexity features of the image, and output noise type labels; Step S12: Based on the noise classification results, different branches are automatically activated: when the noise is Gaussian noise, frequency domain Wiener filtering is performed for enhancement; when the noise is salt-and-pepper noise, nonlocal mean suppression is performed; when the noise is mixed noise, cascaded multi-stage processing is performed, first using nonlocal mean or morphological filtering to eliminate mud contamination, and then using frequency domain Wiener filtering to process dust. Step S13, Training Data and Loss Function: Collect core images under different drilling environments, label noise areas, and use generative adversarial networks to simulate complex noise data. Add MMD loss, fine-tune the generator to reduce the distribution gap between synthesized and real noise, use MS-SSIM to prevent over-smoothing, and use L1 norm to constrain noise residuals and hybrid loss functions.

2. The intelligent identification system for core lithology images according to claim 1, characterized in that: The AI ​​noise reduction is based on a convolutional neural network built on the PyTorch / TensorFlow framework, supports multi-task classification, and introduces residual connections and attention mechanisms.

3. The intelligent identification system for core lithology images according to claim 1, characterized in that: In the feature extraction and classification module, the training steps of the deep learning model include: Step S21, Data preparation and enhancement: Prepare 20,000 to 30,000 images with more than 20 million pixels after denoising and synchronously recorded metadata, and perform data enhancement operations such as small-scale rotation, scaling and / or cropping on the images; Step S22, Design the model architecture: Design a deep learning model using the PyTorch / TensorFlow framework and connect it to a database containing regional geological information. Input the denoised image and metadata to obtain the lithology category probability. Then, stitch the metadata to the fully connected layer for multimodal fusion and use MMD loss to perform domain adaptation training on the core image. Step S23, Combating overfitting: Using spatial stratified sampling, the dataset is divided into training and validation sets based on the drilling location. If the accuracy of the validation set does not significantly improve within several consecutive training cycles, training is terminated. Step S24, Model Validation: Establish a local core feature database, store the defined classification criteria, and associate it with regional stratigraphic information. Use the Torchcam library of PyTorch to generate activation heatmaps of CNN classification results, locate the key areas of interest of the model, match the heatmap features with the standard features in the local feature library, and verify the rationality of the model.

4. The intelligent identification system for core lithology images according to claim 1, characterized in that: The steps for automating the filling of groundwater monitoring well construction records using Python scripts include: Step S31, Data Acquisition and Preprocessing: Extract key parameters from the classification results output by the deep learning model as the basic data source for well construction records, combined with auxiliary data collected on site, including geographic coordinates, time, and borehole depth; Step S32, Standardized Template Design: Design a template for the groundwater monitoring well construction record form, including the basic information area, technical parameter area, hydrogeological data, and graphic area; Step S33: Use Python scripts to automate document generation: automatically fill template fields using a document processing library, match technical parameters according to classification results, and automatically generate groundwater monitoring well construction record forms that conform to industry standards.

5. The intelligent identification system for core lithology images according to claim 4, characterized in that: The basic information area includes: project name, weather conditions, well number, drilling unit, drilling rig model and parameters, drilling parameters, well casing parameters, and well construction time; the technical parameters area includes: filter pipe parameters, filter media type, well cover type, and well bottom sealing method; the hydrogeological data includes: surface elevation, initial water level depth, stable water level depth, and well flushing method.

Images