Deep Semantic Segmentation Research Articles

Seismic horizon tracking based on TransUnet model

Seismic horizons are of paramount importance for building structure models and stratigraphic interpretations. However, currently, horizon tracking is generally obtained through manual tracking or a combination of traditional automatic tracking and manual tracking, which is a highly time-consuming and error-prone process. Despite the many methods proposed for automated horizon tracking, tracking horizons with complex features still face numerous challenges.This paper proposes a method for horizon tracking using the TransUnet model, which can be applied to track both individual horizons and multiple horizons simultaneously. This method is implemented through deep learning semantic segmentation. Training is conducted using a deep learning encoder-decoder structure to establish the mapping relationship between seismic data and horizon data. The TransUnet model builds upon the Unet model by incorporating Transformer modules, enabling the model to possess both local and global attention capabilities. The ultimate results of horizon tracking reveal that the TransUnet model is more proficient at tracking horizons in areas with complex seismic reflections compared to the Unet model. In single and multiple horizon tracking, the root mean square error (RMSE) and coefficient of determination (R2 ) calculated from horizons predicted by the TransUNet model and manually tracked horizons are significantly lower than those of the U-Net model, this research demonstrates that the proposed method is more efficient and accurate in tracking the horizons of three-dimensional seismic volumes compared to traditional approaches.

Submarine Landslide Identification Based on Improved DeepLabv3 with Spatial and Channel Attention

As one of the most destructive, hazardous, and frequent marine geohazards, correctly recognizing submarine landslides holds substantial importance for regional risk assessment, disaster prevention, and marine resource development. Many conventional approaches to prediction and mapping necessitate the involvement of expert insights, oversight, and extensive field investigations, which can result in significant time and effort invested in the prediction process. This paper focuses on employing a deep neural network semantic segmentation technique to detect submarine landslides to replace previous methods, such as numerical analysis and physical modeling, to predict and identify the landslide areas quickly. The peripheral zone of the western Iberian Sea is selected as the study area. Since the neural network image recognition task usually requires RGB images as input data, factors such as slope, hillshade, and elevation extracted from digital elevation model (DEM) data are used to synthesize RGB images through band synthesis methods, and the number and diversity of data are increased utilizing data enhancement. Based on the classical semantic segmentation model DeepLabV3, this paper proposes an improved deep learning method, which strengthens the ability of model feature extraction for complex situations by adding an attention mechanism module, improving the spatial pyramid pooling module, and improving the landslide intersection over union metric from 0.4257 to 0.5219 and the F1-score metric from 0.609 to 0.6631 to achieve effective identification of submarine landslides.

Identification of sweetpotato virus disease-infected leaves from field images using deep learning

IntroductionSweetpotato virus disease (SPVD) is widespread and causes significant economic losses. Current diagnostic methods are either costly or labor-intensive, limiting both efficiency and scalability.MethodsThe segmentation algorithm proposed in this study can rapidly and accurately identify SPVD lesions from field-captured photos of sweetpotato leaves. Two custom datasets, DS-1 and DS-2, are utilized, containing meticulously annotated images of sweetpotato leaves affected by SPVD. DS-1 is used for training, validation, and testing the model, while DS-2 is exclusively employed to validate the model’s reliability. This study employs a deep learning-based semantic segmentation network, DeepLabV3+, integrated with an Attention Pyramid Fusion (APF) module. The APF module combines a channel attention mechanism with multi-scale feature fusion to enhance the model’s performance in disease pixel segmentation. Additionally, a novel data augmentation technique is utilized to improve recognition accuracy in the edge background areas of real large images, addressing issues of poor segmentation precision in these regions. Transfer learning is applied to enhance the model’s generalization capabilities.ResultsThe experimental results indicate that the model, with 62.57M parameters and 253.92 Giga Floating Point Operations Per Second (GFLOPs), achieves a mean Intersection over Union (mIoU) of 94.63% and a mean accuracy (mAcc) of 96.99% on the DS-1 test set, and an mIoU of 78.59% and an mAcc of 79.47% on the DS-2 dataset.DiscussionAblation studies confirm the effectiveness of the proposed data augmentation and APF methods, while comparative experiments demonstrate the model’s superiority across various metrics. The proposed method also exhibits excellent detection results in simulated scenarios. In summary, this study successfully deploys a deep learning framework to segment SPVD lesions from field images of sweetpotato foliage, which will contribute to the rapid and intelligent detection of sweetpotato diseases.

A Comparison of Uncertainty Estimation Methods for Building Footprint Change Detection from Sentinel-2 Imagery

Abstract. This manuscript investigates the effects of uncertainty methods applied to the problem of deep learning-based semantic segmentation of building footprints on moderate-resolution satellite imagery. While the recent efforts of big corporations to add information about building locations and sizes on a global or continental scale are generally valuable, still the overall challenge persists in identifying the spatial-temporal patterns of growing urbanization. In this work, we extend UNet-type architectures to perform binary building footprint classification based on Sentinel-2 imagery resulting in five different models. While previous studies focused on urban areas in the Western world we conduct all training and evaluation in India. All models are trained on Microsoft building footprint products while for evaluation purposes high-quality reference data is manually selected from regions with especially good open-street-map coverage. Quantitative and qualitative experiments are conducted where a significant performance gain is found for a model trained with a mixture of aleatoric and epistemic uncertainty measures. The performance gain is even more pronounced for subsequent quantitative multi-temporal change detection experiments.

Crop Classification from Drone Imagery Based on Lightweight Semantic Segmentation Methods

Technological advances have dramatically improved precision agriculture, and accurate crop classification is a key aspect of precision agriculture (PA). The flexibility and real-time nature of UAVs have led them to become an important tool for acquiring agricultural data and enabling precise crop classification. Currently, crop identification relies heavily on complex high-precision models that often struggle to provide real-time performance. Research on lightweight models specifically for crop classification is also limited. In this paper, we propose a crop classification method based on UAV visible-light images based on PP-LiteSeg, a lightweight model proposed by Baidu. To improve the accuracy, a pyramid pooling module is designed in this paper, which integrates adaptive mean pooling and CSPC (Convolutional Spatial Pyramid Pooling) techniques to handle high-resolution features. In addition, a sparse self-attention mechanism is employed to help the model pay more attention to locally important semantic regions in the image. The combination of adaptive average pooling and the sparse self-attention mechanism can better handle different levels of contextual information. To train the model, a new dataset based on UAV visible-light images including nine categories such as rice, soybean, red bean, wheat, corn, poplar, etc., with a time span of two years was created for accurate crop classification. The experimental results show that the improved model outperforms other models in terms of accuracy and prediction performance, with a MIoU (mean intersection ratio joint) of 94.79%, which is 2.79% better than the original model. Based on the UAV RGB images demonstrated in this paper, the improved model achieves a better balance between real-time performance and accuracy. In conclusion, the method effectively utilizes UAV RGB data and lightweight deep semantic segmentation models to provide valuable insights for crop classification and UAV field monitoring.

GPS-free autonomous navigation in cluttered tree rows with deep semantic segmentation

Segmentation-based autonomous navigation has recently been presented as an appealing approach to guiding robotic platforms through crop rows without requiring perfect GPS localization. Nevertheless, current techniques are restricted to situations where the distinct separation between the plants and the sky allows for the identification of the row’s center. However, tall, dense vegetation, such as high tree rows and orchards, is the primary cause of GPS signal blockage. In this study, we increase the overall robustness and adaptability of the control algorithm by extending the segmentation-based robotic guiding to those cases where canopies and branches occlude the sky and prevent the utilization of GPS and earlier approaches. An efficient Deep Neural Network architecture has been used to address semantic segmentation, performing the training with synthetic data only. Numerous vineyards and tree fields have undergone extensive testing in both simulation and real world to show the solution’s competitive benefits. The system achieved unseen results in orchards, with a Mean Average Error smaller than 9% of the maximum width of each row, improving state-of-the-art algorithms by disclosing new scenarios such as close canopy crops. The official code can be found at: https://github.com/PIC4SeR/SegMinNavigation.git.

Digital rock modeling of deformed multi-scale media in deep hydrocarbon reservoirs based on in-situ stress-loading CT imaging and U-Net deep learning

Deep and ultra-deep hydrocarbon reservoirs are regarded as an attractive target for onshore oil and gas exploration and development in China. The multi-scale fractures are widely distributed in this type of reservoir. The fluid-solid coupling effect is strong, which can greatly affect the fluid flow, resulting in a low hydrocarbon recovery with the average value less than 15%. To explore the underlying flow dynamics during efficient development of deep and ultra-deep hydrocarbon reservoirs, it is of great importance to characterize the multi-scale fracture-pore structure evolution of rock affected by strong geo-stress field variation. To tackle the above issues, an actual rock drilled from a typical ultra-deep reservoir in Tarim Basin are selected to conduct in-situ stress-loading computed tomography (CT) scanning experiments, and the CT gray images of rock microstructure under different dynamic loading conditions are obtained. A fully convolutional neural network (U-Net) deep learning segmentation algorithm is introduced to accurately distinguish the rock skeleton, pore space and fractures in deep rock. The three-dimensional (3-D) digital rock model of deformed multi-scale fracture-porous media under different dynamic loading conditions is ultimately established to investigate the evolution of fracture morphology and deep rock microstructure as the in-situ stress is gradually loaded. It indicates that, the U-Net deep learning semantic segmentation algorithm can accurately segment CT images of deep rocks, and the established 3D digital rock model of fractured porous media can accurately represent the pore-throat distribution, fracture morphology, and pore-scale topological connectivity. At the initial stage of stress loading, there are few micro-fractures inside the rock, and the fracture connectivity is poor. Due to effect of rock compaction, the average pore throat radius increases while pore throat ratio, coordination number and tortuosity greatly decrease. As the effective stress is gradually loaded, the micro-fractures begin to propagate, leading to stronger heterogeneity of micro-fractures’ distribution and better topological connectivity, and both the coordination number and tortuosity gradually increase. After the deep rock is fractured, micro-fractures run through a fracture network and all the above topological parameters increase greatly.

Improved lane detection for autonomous vehicles using deep learning, semantic segmentation, edge detection and multi-sensor data fusion

Automated driving has gained significant attention because it can eliminate severe driving risks in real time. While autonomous vehicles rely heavily on sensors for lane detection, obstacle identification, and environmental awareness, accurate lane recognition remains a persistent challenge due to factors such as noise from shadows, poor lane markings, and obstructed views. Despite advances in computer vision, this problem is yet to be fully resolved, presenting a gap in the current literature. The primary objective of this research is to address these challenges by developing an enhanced lane-detection system. To achieve this, the study integrates advanced techniques, including semantic segmentation, edge detection, and deep learning, coupled with multi-sensor data fusion from cameras, LIDAR, and radar. By employing this methodology, the research examines various lane-detection methods and benchmarks the proposed model against existing systems in terms of accuracy, specificity and processing speed. Initial findings demonstrate that the combination of semantic segmentation and multi-sensor fusion improves lane detection in real-time scenarios. The proposed model achieved a lane detection accuracy of 97.8%, a specificity of 99.28%, and an average processing time of 0.0047 seconds per epoch. This study not only addresses the limitations of existing lane detection systems but also offers insights into improving road safety for autonomous vehicles.

LSL-SS-Net: level set loss-guided semantic segmentation networks for landslide extraction

ABSTRACT A landslide inventory is essential for numerous applications, including disaster prevention and mitigation. The emergence of advanced satellite technologies and the proliferation of extensive satellite imagery have greatly enhanced the role of deep learning-based semantic segmentation networks in the development of comprehensive landslide inventories. In deep learning, especially in semantic segmentation tasks, the loss function is crucial to guide the training of neural networks. However, typical loss functions have several challenges and shortcomings, including difficulty in handling objects of different scales and neglecting spatial correlation. These drawbacks hinder semantic segmentation networks from solely concentrating on accurately predicting landslide categories with rich semantic information. The spatial details (e.g. small landslides and precise landslide boundary information) in the segmentation results may be lost, reducing the accuracy of landslide extraction. Most studies only focus on training and detection in the local region. Consequently, detecting new unexplored region and detrimental to emergency response is challenging. This study presents a level set loss-guided semantic segmentation network, which can be integrated with different semantic segmentation networks, to address these issues. Three out of five study areas (e.g. Bijie, Jiuzhaigou, and Taitung) are used for model training and accuracy evaluation, while the rest, which are newly unexplored regions (e.g. Hokkaido and Haiti), are used to test the transferability of the proposed loss function in different regions. The study results validate the effectiveness of the designed level set loss in various semantic segmentation networks and in different study areas for landslide extraction.

Urban Color Perception and Sentiment Analysis Based on Deep Learning and Street View Big Data

The acceleration of urbanization has resulted in a heightened awareness of the impacts of urban environments on residents’ emotional states. This present study focuses on the Lixia District of Jinan City. By using urban street view big data and deep learning methods, we undertook a detailed analysis of the impacts of urban color features on residents’ emotional perceptions. In particular, a substantial corpus of street scene image data was extracted and processed. This was performed using a deep convolutional neural network (DCNN) and semantic segmentation technology (PSPNet), which enabled the simulation and prediction of the subjective perception of the urban environment by humans. Furthermore, the color complexity and coordination in the street scene were quantified and combined with residents’ emotional feedback to carry out a multi-dimensional analysis. The findings revealed that color complexity and coordination were significant elements influencing residents’ emotional perceptions. A high color complexity is visually appealing, but can lead to fatigue, discomfort, and boredom; a moderate complexity stimulates vitality and pleasure; high levels of regional harmony and aesthetics can increase perceptions of beauty and security; and low levels of coordination can increase feelings of depression. The environmental characteristics of different areas and differences in the daily activities of residents resulted in regional differences regarding the impacts of color features on emotional perception. This study corroborates the assertion that environmental color coordination has the capacity to enhance residents’ emotions, thereby providing an important reference point for urban planning. Planning should be based on the functional characteristics of the region, and color complexity and coordination should be reasonably regulated to optimize the emotional experiences of residents. Differentiated color management enhances urban aesthetics, livability, and residents’ happiness and promotes sustainable development. In the future, the influences of color and environmental factors on emotions can be explored in depth, with a view to assist in the formulation of fine urban design.

A Modified Deep Semantic Segmentation Model for Analysis of Whole Slide Skin Images

Automated segmentation of biomedical image has been recognized as an important step in computer-aided diagnosis systems for detection of abnormalities. Despite its importance, the segmentation process remains an open challenge due to variations in color, texture, shape diversity and boundaries. Semantic segmentation often requires deeper neural networks to achieve higher accuracy, making the segmentation model more complex and slower. Due to the need to process a large number of biomedical images, more efficient and cheaper image processing techniques for accurate segmentation are needed. In this article, we present a modified deep semantic segmentation model that utilizes the backbone of EfficientNet-B3 along with UNet for reliable segmentation. We trained our model on Non-melanoma skin cancer segmentation for histopathology dataset to divide the image in 12 different classes for segmentation. Our method outperforms the existing literature with an increase in average class accuracy from 79 to 83%. Our approach also shows an increase in overall accuracy from 85 to 94%.

A machine vision‐based intelligent segmentation method for dam underwater cracks using swarm optimization algorithm and deep learning

AbstractEnsuring the safety of water networks is a research hotspot in the current water conservancy industry, and dams are an important part. However, over time, the dam is prone to varying degrees of aging and disease, most of which are structural cracks. If they cannot be discovered and repaired in time, the normal operation of the dam will be affected, and even catastrophic accidents such as dam failure will occur. However, complex backgrounds and blurred images can easily lead to misjudgments by machine vision detection models, and high‐efficiency and accurate detection and evaluation technology are urgently needed. This paper combines the deep semantic segmentation network and the model hyperparameters optimization algorithm to propose a data‐intelligent perception method of dam underwater cracks driven by knowledge coupling. Taking the underwater detection of a concrete face rockfill dam as an example, the effectiveness of the model is verified by using the underwater vehicle as the carrier. Experimental results indicate that the developed method achieves an intersection‐union ratio of 0.9301, a precision rate of 0.9678, a precision rate of 0.9472, and a recall rate of 0.9577 in the test set. This shows that the constructed method has a high crack fine detection performance. In addition, the developed method has better segmentation performance in different complex underwater crack scenes, which further illustrates the high performance of the developed method.

Lake responses and mechanisms to El Niño on the Tibetan Plateau using deep learning-based semantic segmentation

Numerous lakes across the Tibetan Plateau (TP) serve as crucial indicators of climate change and are significantly influenced by El Niño events. Previous studies of lake response to El Niño events have focused on a limited number of lakes. Despite advances in remote sensing technology, there have been few comprehensive and large-scale studies using deep learning, and there are still gaps in understanding the response mechanisms on a larger scale. This study leverages advanced deep-learning techniques to map lake responses, offering unprecedented insights into the large-scale hydrological impacts of El Niño. Our results show that lakes shrink significantly during El Niño events on the TP. Lakes located in the central and southern parts of the TP and small lakes with areas ranging from 1 to 50 km2 (over 60 % of them) exhibited strong responses. The range of lake response to El Niño events varies with their intensity, with stronger El Niño events causing an expansion of the response range along the latitudinal direction. We propose four possible mechanisms for lake response patterns to El Niño from the perspective of lake water sources. Strong shrinkage is primarily caused by decreased precipitation and increased evaporation, with a possible contribution from reduced meltwater. Strong expansion is due to increased precipitation, more glacier and frozen soil meltwater, and reduced evaporation. For slight shrinkage and expansion patterns, the balance of meltwater may offset or even counteract the El Niño signal. The study’s results could improve predictions of extreme weather events like droughts and floods in the Third Pole region, enhance water resource management and responsiveness, and offer valuable insights for ecological monitoring and early warning systems development.

Real-time segmentation of biliary structure in pure laparoscopic donor hepatectomy

Pure laparoscopic donor hepatectomy (PLDH) has become a standard practice for living donor liver transplantation in expert centers. Accurate understanding of biliary structures is crucial during PLDH to minimize the risk of complications. This study aims to develop a deep learning-based segmentation model for real-time identification of biliary structures, assisting surgeons in determining the optimal transection site during PLDH. A single-institution retrospective feasibility analysis was conducted on 30 intraoperative videos of PLDH. All videos were selected for their use of the indocyanine green near-infrared fluorescence technique to identify biliary structure. From the analysis, 10 representative frames were extracted from each video specifically during the bile duct division phase, resulting in 300 frames. These frames underwent pixel-wise annotation to identify biliary structures and the transection site. A segmentation task was then performed using a DeepLabV3+ algorithm, equipped with a ResNet50 encoder, focusing on the bile duct (BD) and anterior wall (AW) for transection. The model’s performance was evaluated using the dice similarity coefficient (DSC). The model predicted biliary structures with a mean DSC of 0.728 ± 0.01 for BD and 0.429 ± 0.06 for AW. Inference was performed at a speed of 15.3 frames per second, demonstrating the feasibility of real-time recognition of anatomical structures during surgery. The deep learning-based semantic segmentation model exhibited promising performance in identifying biliary structures during PLDH. Future studies should focus on validating the clinical utility and generalizability of the model and comparing its efficacy with current gold standard practices to better evaluate its potential clinical applications.

Deep handwritten diagram segmentation

AbstractHandwriting is a natural way to communicate and exchange ideas, but converting handwritten diagrams to application‐specific digital formats requires skill and time. Automatic handwritten document conversion can save time, but diagrams and text require different recognition engines. Since accurate segmentation of handwritten diagrams can improve the accuracy of later diagram recognition steps, the authors propose to solve the problem of segmentation of text and non‐text elements of handwritten diagrams using deep semantic segmentation. The model, DeepDP is a flexible U‐net style architecture that can be tuned in complexity to a level appropriate for a particular dataset and diagram type. Experiments on a public hand‐drawn flowchart dataset and a business process diagram dataset show excellent performance, with a per pixel accuracy of 98.6% on the public flowchart datasets and improvement over the 99.3% text stroke accuracy and 96.6% non‐text stroke accuracy obtained by state of the art methods that use online stroke information. On the smaller offline business process diagram dataset, the method obtains a per‐pixel accuracy of 96.9%.

Semantic Segmentation and Classification of Active and Abandoned Agricultural Fields through Deep Learning in the Southern Peruvian Andes

The monumental scale agricultural infrastructure systems built by Andean peoples during pre-Hispanic times have enabled intensive agriculture in the high-relief, arid/semi-arid landscape of the Southern Peruvian Andes. Large tracts of these labor-intensive systems have been abandoned, however, owing in large measure to a range of demographic, economic, and political crises precipitated by the Spanish invasion of the 16th century CE. This research seeks to better understand the dynamics of agricultural intensification and deintensification in the Andes by inventorying through the semantic segmentation of active and abandoned agricultural fields in satellite imagery across approximately 77,000 km2 of the Southern Peruvian Highlands. While manual digitization of agricultural fields in satellite imagery is time-consuming and labor-intensive, deep learning-based semantic segmentation makes it possible to map and classify en masse Andean agricultural infrastructure. Using high resolution satellite imagery, training and validation data were manually produced in distributed sample areas and were used to transfer-train a convolutional neural network for semantic segmentation. The resulting dataset was compared to manual surveys of the region and results suggest that deep learning can generate larger and more accurate datasets than those generated by hand.

Automatic generation of diffusion tensor imaging for the lumbar nerve using convolutional neural networks

【Purpose】Diffusion Tensor Imaging (DTI) with tractography is useful for the functional diagnosis of degenerative lumbar disorders. However, it is not widely used in clinical settings due to time and health care provider costs, as it is performed manually on hospital workstations. The purpose of this study is to construct a system that extracts the lumbar nerve and generates tractography automatically using deep learning semantic segmentation. 【Methods】We acquired 839 axial diffusion weighted images (DWI) from the DTI data of 90 patients with degenerative lumbar disorders, and segmented the lumbar nerve roots using U-Net, a semantic segmentation model. Using five architectural models, the accuracy of the lumbar nerve root segmentation was evaluated using a Dice coefficient. We also created automatic scripts from three commercially available software tools, including MRICronGL for medical image viewing, Diffusion Toolkit for reconstruction of the DWI data, and Trackvis for the creation of the tractography, and compared the time required to create the tractography, and evaluated the quality of the automated tractography was evaluated. 【Results】Among the five models, the architectural model Resnet34 performed the best with a Dice = 0.780. The creation time for the automatic lumbar nerve tractography was 191 s, which was significantly shorter by 235 s than the manual time of 426 s (p < 0.05). Furthermore, the agreement between manual and automated tractography was 3.67 ± 1.53 (satisfactory). 【Conclusions】Using deep learning semantic segmentation, we were able to construct a system that automatically extracted the lumbar nerve and generated lumbar nerve tractography. This technology makes it possible to analyze lumbar nerve DTI and create tractography automatically, and is expected to advance the clinical applications of DTI for the assessment of the lumbar nerve.

Innovative: A Novel Deep Learning-Based Semantic Segmentation Architecture for Medical Applications

Innovative: A Novel Deep Learning-Based Semantic Segmentation Architecture for Medical Applications

Automatic topology and capacity generation framework for urban drainage systems with deep learning-based land use segmentation and hydrological characterization

This paper proposed a comprehensive and automated framework that integrates the entire urban drainage design process, from refined land use segmentation, to catchment subdivision and characterization, to network topology generation, and finally to hydraulic estimation of pipeline capacity, based on multiple open-source datasets. The method incorporated deep learning-based semantic segmentation, GIS-based hydrological characterization, topology generation and hydraulic calculation algorithms. The performance and applicability of the method were validated through case studies of two cities. The results showed that the deep learning model with prior knowledge could accurately segment urban land use to obtain the detailed shape and boundary description, and achieved higher detection accuracy for water bodies, buildings, and roads. The stormwater drainage topology and conveyance capacity were automatically generated based on subcatchment division and hydrological characterization under a set of topological parameter conditions. In the optimal Critical Success Index (CSI) scenario, only 8% and 21% of the generated pipelines were inconsistent with the expert-designed networks. Compared to the traditional manual design process, the automated approach can save significant time and resources and reduce the reliance on field surveys and empirical/expert work. At the same time, the tool can be employed for system layout and capacity optimization, providing important support for the automation, design, and rehabilitation of stormwater drainage systems.

Point cloud semantic segmentation of grotto scenes using the knowledge-guided deep learning method

ABSTRACT In recent years, the deep learning-based semantic segmentation for point clouds has demonstrated remarkable capabilities in processing 3D urban scenes for applications such as three-dimensional reconstruction, semantic modeling, and augmented reality. However, research on grottoes scenes is very limited. It is currently unclear how existing neural architectures for point cloud semantic segmentation perform in grotto scenes, and how to effectively incorporate the unique characteristics of grotto scenes to enhance the performance of deep neural networks. This study proposed a method for point cloud semantic segmentation of grotto scenes, combining knowledge with deep learning approaches. The method adopted knowledge to guide the creation of benchmark datasets, the design of a neural network called GSS-Net, and the correction of segmentation errors in the results of deep learning. The results show that the proposed method outperforms four existing mainstream models without the correction of segmentation results. Moreover, a set of ablation studies verified the effectiveness of each proposed module. This method not only improves the accuracy of point cloud semantic segmentation in grotto scenes but also enhances the interpretability of network designs. It provides new insights into the application of knowledge-guided deep learning models in grotto scenes.