Remote sensing segmentation method and system for pegmatite vein of deep cutting area fused with three-dimensional terrain

Abstract

The application discloses a kind of deep cutting area pegmatite vein remote sensing segmentation methods and systems of fusing three-dimensional terrain, and relates to geological remote sensing technical field.For the problem that prior art does not fully consider three-dimensional information of deep cutting complex terrain, resulting in low pegmatite vein remote sensing segmentation precision and poor spatial continuity, the application obtains multispectral remote sensing image and laser radar digital elevation model data of the study area, expands the two-dimensional space-spectral characteristics of the pixel into three-dimensional characteristics through a three-dimensional terrain fusion module, builds a space-spectral Mamba model that fuses three-dimensional terrain, and uses a loss function containing an elevation constraint term to complete model training, achieving accurate remote sensing segmentation of pegmatite veins.The application significantly improves the vein segmentation accuracy and spatial continuity under complex terrain, and is suitable for exploration of complex mountainous granite pegmatite type lithium mine.

Description

Technical Field

[0001] This invention relates to the field of geological remote sensing technology, and in particular to a remote sensing segmentation method and system for pegmatite veins in deeply cut areas that integrates three-dimensional topography. Background Technology

[0002] As a strategic mineral resource in China, lithium is an important source of lithium resources, and granite pegmatite-type lithium deposits are the core geological bodies that host this type of lithium deposit. The accurate identification and delineation of its spatial location, distribution morphology and continuity directly determines the reliability of lithium resource surveys, reserve estimation and prospecting target area prediction. It is a core technical link in lithium exploration in complex mountainous areas.

[0003] Remote sensing technology, with its advantages of large-scale, non-contact, and rapid imaging, has become the mainstream technology for geological body identification and mineral exploration. High-resolution multispectral remote sensing can efficiently acquire surface spectral and planar spatial features, supporting the detailed interpretation of geological bodies. Based on this, traditional machine learning (support vector machines, K-nearest neighbors, etc.) and deep learning (two-dimensional convolutional neural networks, three-dimensional convolutional neural networks, spatial-spectral Transformers, etc.) segmentation methods have been successively applied to the remote sensing identification of pegmatite veins, gradually realizing the technological iteration from manual interpretation to automated and intelligent segmentation, providing basic technical support for lithium remote sensing exploration.

[0004] However, in deeply dissected and complex mountainous areas with dramatic topographic relief, numerous valleys, and varying slopes and aspects, existing remote sensing segmentation methods have significant technical limitations: current models are all based on two-dimensional remote sensing images, utilizing only pixel spectral features and planar spatial relationships for modeling, without incorporating three-dimensional topographic information. Dramatic topographic relief can cause pixel shifts in remote sensing images and distortions in coverage, leading to spectral heterogeneity and spatial representation breaks within the same pegmatite vein; simultaneously, traditional models lack vertical topographic constraints, making it impossible to adapt to the three-dimensional distribution of veins along topographic relief, easily resulting in problems such as missing vein segments, boundary shifts, and poor spatial continuity. The segmentation accuracy and practicality are insufficient to meet the engineering requirements of lithium exploration in complex mountainous areas.

[0005] Currently, the industry has not yet developed a dedicated remote sensing segmentation technology for pegmatite veins that integrates three-dimensional terrain information and is adapted to deeply dissected landforms. There is a significant technical gap in the high-precision and high-continuity segmentation of mineral veins under complex terrain. Developing a new remote sensing segmentation method for pegmatite veins adapted to deeply dissected mountainous areas has become a key technical problem that urgently needs to be solved in the field of lithium mine remote sensing exploration. Summary of the Invention

[0006] The purpose of this invention is to provide a remote sensing segmentation method and system for pegmatite veins in deeply incised areas that integrates three-dimensional terrain, in order to solve the technical problems of discontinuous pegmatite vein segmentation, boundary offset, and low identification accuracy caused by traditional two-dimensional remote sensing segmentation methods under deeply incised complex terrain conditions due to neglecting the influence of terrain undulation.

[0007] To achieve the above objectives, this invention provides a remote sensing segmentation method for pegmatite veins in deeply dissected areas that incorporates three-dimensional terrain data. The steps are as follows:

[0008] S1. Acquire multispectral remote sensing images and lidar point cloud data of the study area, preprocess the multispectral remote sensing images, process the lidar point cloud data to generate a digital elevation model, and spatially register the digital elevation model with the preprocessed multispectral remote sensing images.

[0009] S2. Construct a three-dimensional terrain information fusion module, introduce the elevation information in the digital elevation model as an independent dimension into the pixel feature space of the multispectral remote sensing image, and expand the original two-dimensional spatial-spectral features of each pixel into a three-dimensional feature representation that includes planar coordinates, elevation and spectral features.

[0010] S3. Construct a spatial-spectral Mamba model that integrates three-dimensional terrain information. Input the three-dimensional feature representation obtained in step S2 into the spatial-spectral Mamba model to complete feature encoding and fusion. Use a total loss function that includes an elevation constraint term to supervise the training of the model.

[0011] S4. Use the trained spatial-spectral Mamba model to segment and identify pegmatite veins in the study area, and output the remote sensing segmentation results of the pegmatite veins.

[0012] Preferably, in step S2, the three-dimensional feature representation specifically includes:

[0013] Let the spectral influence be defined in a two-dimensional spatial domain. The corresponding digital elevation model is a continuous elevation function. ,in, Indicates pixel position The surface elevation value at the location; each pixel consists of the original two-dimensional spatial-spectral features. Extended to three-dimensional feature representation , where s represents the spectral feature vector of the corresponding pixel.

[0014] Preferably, in step S3, the space-spectral Mamba model includes a spectral feature extraction branch, a spatial feature extraction branch, and multiple cascaded space-spectral Mamba modules; in the model input stage, the elevation information of the digital elevation model is embedded as an independent channel into the spatial feature token, and the information of each band of the multispectral image is constructed into a spectral token, which is then input into the space-spectral Mamba encoder to complete feature extraction.

[0015] Preferably, each of the spatial-spectral Mamba modules includes a linear mapping layer, a state space model unit, and a normalization structure. The state space model unit is used to model long-distance dependencies of the input sequence to achieve feature fusion of spatial and spectral dimensions.

[0016] Preferably, in step S3, the total loss function is defined as:

[0017]

[0018] in, The supervised segmentation loss function is based on labeled samples. This is an elevation constraint loss term, used to incorporate terrain information to constrain the spatial continuity of the segmentation results. These are terrain weight parameters.

[0019] Preferably, in step S3, the model training adopts a dataset construction strategy based on spatial region division: the study area is divided into several independent spatial regions, one part of which is used as the training set and the remaining part is used as the validation set, with no spatial overlap between the training set and the validation set.

[0020] Preferably, in step S1, the multispectral remote sensing image is a high-resolution multispectral satellite image, and the preprocessing includes radiometric calibration and atmospheric correction; the digital elevation model is generated from point cloud data acquired by UAV-borne or airborne lidar systems through trajectory calculation, filtering, and ground point extraction, and the digital elevation model is resampled to the same spatial resolution as the multispectral remote sensing image.

[0021] Preferably, in step S4, after obtaining the remote sensing segmentation results of the pegmatite veins, the overall accuracy OA, average accuracy AA, and Kappa coefficient are used as evaluation indicators to quantitatively evaluate the accuracy of the model segmentation results; field observations are conducted in the pegmatite vein area within the study area, rock samples are collected for thin section microscopy analysis, and the field verification of the segmentation results is completed.

[0022] A remote sensing segmentation system for pegmatite veins in deeply dissected areas, incorporating three-dimensional terrain, is used to implement the remote sensing segmentation method for pegmatite veins in deeply dissected areas, as described above, comprising:

[0023] The data acquisition and preprocessing module is used to acquire multispectral remote sensing images and lidar point cloud data of the study area, preprocess the multispectral remote sensing images, process the lidar point cloud data to generate a digital elevation model, and spatially register the digital elevation model with the preprocessed multispectral remote sensing images.

[0024] The three-dimensional terrain fusion module is used to introduce the elevation information in the digital elevation model as an independent dimension into the pixel feature space of the multispectral remote sensing image, and to expand the original two-dimensional spatial-spectral features of each pixel into a three-dimensional feature representation that includes planar coordinates, elevation and spectral features.

[0025] The model building and training module is used to build a spatial-spectral Mamba model that integrates three-dimensional terrain information. The three-dimensional feature representation is input into the spatial-spectral Mamba model to complete feature encoding and fusion. The model is trained under supervision using a total loss function that includes an elevation constraint term.

[0026] The segmentation and recognition module is used to segment and recognize pegmatite veins in the study area using a trained spatial-spectral Mamba model, and output the remote sensing segmentation results of the pegmatite veins.

[0027] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the remote sensing segmentation method for deeply dissected pegmatite veins incorporating three-dimensional terrain as described above.

[0028] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0029] 1. This invention breaks through the core technical bottleneck of traditional two-dimensional remote sensing segmentation methods. It addresses the inherent defects of existing technologies that rely solely on planar spectral and spatial feature modeling and do not fully consider the vertical variations of deeply cut terrain. By introducing DEM elevation information as an independent dimension into the pixel feature space, a three-dimensional spatial-spectral-terrain joint feature expression is constructed. This enables the model to simultaneously perceive the spatial distribution patterns of pegmatite veins in both the planar and vertical directions, fundamentally solving the industry pain point of spectral heterogeneity and spatial expression fragmentation of the same geological body caused by terrain undulation.

[0030] 2. This invention introduces terrain spatial constraints throughout the entire model training and inference process by using a terrain embedding module and a total loss function with elevation constraints. This effectively overcomes the problems of segmentation, target loss, and boundary offset in areas with drastic terrain changes such as steep slopes, valleys, and slope breaks in deeply cut mountainous areas. It significantly improves the spatial continuity and boundary integrity of the pegmatite vein segmentation results, enabling the segmentation results to accurately match the real spatial distribution of pegmatite veins in complex terrain.

[0031] 3. This invention relies on the long-distance dependency efficiency modeling capability of the spatial-spectral Mamba model and combines it with three-dimensional terrain information fusion to achieve multi-dimensional feature collaborative modeling. While reducing the computational complexity of the model, it significantly enhances the model's ability to identify fine-grained features of narrow strip-shaped pegmatite veins. It effectively improves the model's anti-interference ability and scene generalization performance under complex terrain and variable surface cover conditions. Compared with traditional machine learning and mainstream deep learning segmentation methods, it has significant performance advantages.

[0032] 4. This invention significantly expands the engineering application value of remote sensing technology in lithium exploration in complex mountainous areas. The method is adapted to the exploration needs of complex geomorphic environments such as deeply dissected mountains. It can achieve efficient and non-contact delineation of the spatial distribution of pegmatite veins without relying on a large amount of ground measurement data, which significantly reduces the field operation cost and safety risks of mineral exploration in complex mountainous areas. It provides stable and reliable technical support for resource investigation and prospecting target area prediction of granite pegmatite-type lithium deposits, and has good engineering applicability and industry promotion value.

[0033] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is an overall flowchart of the remote sensing segmentation method for pegmatite veins in deeply cut areas that incorporates three-dimensional terrain, according to an embodiment of the present invention.

[0036] Figure 2 This is a schematic diagram of the multi-source data acquisition and preprocessing process in the study area according to an embodiment of the present invention, wherein (a) is a schematic diagram of multispectral image acquisition, (b) is a schematic diagram of terrain data acquisition, (c) is multispectral image data, and (d) is high-precision terrain data;

[0037] Figure 3 This is a schematic diagram comparing two-dimensional and three-dimensional segmentation of pegmatite veins according to an embodiment of the present invention, wherein (a) is a three-dimensional topography of the study area, (b) is a schematic diagram of three-dimensional topography, (c) is a schematic diagram of traditional two-dimensional vein segmentation, and (d) is a schematic diagram of three-dimensional vein segmentation.

[0038] Figure 4 This is a schematic diagram of the structure of the three-dimensional terrain information fusion module in an embodiment of the present invention, wherein (a) is the process of fusion of multispectral image and terrain information, and (b) is a schematic diagram of fusion of spectrum and terrain.

[0039] Figure 5 This is a schematic diagram of the spatial-spectral Mamba model structure that integrates three-dimensional terrain information according to an embodiment of the present invention;

[0040] Figure 6 This is a confusion matrix diagram of the model segmentation results in an embodiment of the present invention;

[0041] Figure 7 This is a three-dimensional segmentation result image of four types of land features in the study area according to an embodiment of the present invention;

[0042] Figure 8 The images show a comparison of two-dimensional and three-dimensional segmentation results of pegmatite veins in deeply cut areas according to an embodiment of the present invention, wherein (a) is an overview of the two-dimensional segmentation results, (b) is an overview of the three-dimensional segmentation results, (c) is a magnified view of a part of the two-dimensional segmentation results, and (d) is a magnified view of a part of the three-dimensional segmentation results.

[0043] Figure 9 These are comparison images of the field verification results of pegmatite vein segmentation in an embodiment of the present invention, where (a) is the pegmatite vein segmentation result of verification area 1, (b) is the field verification result of pegmatite vein in verification area 1, (c) is the pegmatite vein segmentation result of verification area 2, and (d) is the field verification result of pegmatite vein in verification area 2. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] Example

[0047] This embodiment provides a remote sensing segmentation method for pegmatite veins in deeply dissected areas, incorporating three-dimensional topographic information. The study area is the Xichangou lithium deposit on the northern margin of the Altun Orogenic Belt in Xinjiang Uygur Autonomous Region. This region features dramatic topographic relief and intense valley incision, representing a typical deeply dissected mountainous landform. Granite pegmatite veins are widely developed within the mining area, exhibiting a banded distribution that is largely consistent with the attitude of the surrounding strata. They are primarily platy or vein-like in shape, representing a typical pegmatite-type lithium deposit occurrence. The complex topographic conditions and relatively good rock exposure provide excellent experimental conditions for conducting remote sensing segmentation and identification research on pegmatite veins under complex topographic conditions.

[0048] like Figure 1 As shown, the specific implementation steps are as follows:

[0049] S1. Acquisition and Preprocessing of Multi-Source Remote Sensing Data for the Study Area

[0050] 1. Data Source Selection: WorldView-3 satellite multispectral imagery was used as the primary optical data source. This satellite imagery possesses high spatial resolution and multi-band spectral information, effectively characterizing surface lithological differences and providing a reliable data foundation for pegmatite vein segmentation and identification. Before use, the original imagery underwent radiometric calibration and atmospheric correction to obtain surface reflectance data.

[0051] To obtain detailed topographic information of the study area, an unmanned aerial vehicle (UAV) equipped with a lidar system was used for aerial surveying to acquire high-precision point cloud data. This UAV can rapidly acquire large-scale topographic data in complex mountainous environments. The lidar system has a maximum range of approximately 1400m, effectively adapting to the terrain conditions of deeply dissected mountainous areas with large elevation differences, enabling high-precision three-dimensional topographic measurement of complex landforms and providing strong data support for subsequent applications.

[0052] 2. Data Preprocessing: By performing trajectory calculation, filtering, and ground point extraction on the point cloud data, a digital elevation model (DEM) is generated. Subsequently, the DEM data is resampled to the same spatial resolution as the multispectral image and spatially registered with the remote sensing image, thereby achieving unified processing of multi-source remote sensing data and providing basic data for subsequent fusion of 3D terrain information and spectral information. Figure 2 The study demonstrates the data acquisition and preprocessing process for the study area.

[0053] S2, Construction of 3D Terrain Information Fusion Module

[0054] Three-dimensional terrain information was incorporated into a multispectral remote sensing image segmentation model to construct a three-dimensional terrain fusion module. For example... Figure 3 As shown, traditional segmentation methods based on two-dimensional remote sensing images usually treat the image as a set of planar pixels and only use the spectral features of pixels and the two-dimensional spatial neighborhood relationship for target segmentation and recognition. In areas with large topographic relief, the methods are easily affected by changes in slope orientation and differences in observation conditions, resulting in discontinuities in the spatial representation of the same geological body in the image, thereby reducing the accuracy of the vein segmentation results.

[0055] A digital elevation model (DEM) generated from lidar data serves as the source of three-dimensional terrain information. The DEM provides a stable and unique surface elevation value for each pixel in the study area, describing the three-dimensional spatial morphology of the surface. By introducing the elevation information from the DEM into the feature space of the multispectral image, spectral and terrain information are fused, thereby expanding the original two-dimensional spectral-spatial features into a three-dimensional feature representation that includes vertical information, such as... Figure 4As shown. This method enables the model to not only utilize the spectral and spatial distribution characteristics of pegmatite veins in the planar direction when segmenting and identifying them, but also to combine topographic elevation change information to constrain the continuity of pegmatite veins in the vertical direction, thereby improving the accuracy and spatial continuity of vein segmentation and identification under complex and deeply dissected terrain conditions.

[0056] The specific process for importing 3D elevation data is as follows:

[0057] Let Worldview3 multispectral imagery be defined in a two-dimensional spatial domain. The corresponding DEM is a continuous elevation function:

[0058] (1)

[0059] in, Indicates pixel position The surface elevation value at that location. By incorporating elevation information into the pixel feature space, each pixel changes from its original two-dimensional spatial-spectral features:

[0060] (2)

[0061] Expanded to a three-dimensional feature representation that includes terrain dimensions:

[0062] (3)

[0063] in, This represents the spectral feature vector of the corresponding pixel. This representation allows the model to simultaneously perceive spectral differences and vertical elevation changes when learning vein segmentation boundaries, thus better reflecting the true spatial distribution of veins under deeply incised terrain conditions.

[0064] S3. Construction and Training of a Spatial-Spectral Mamba Model Integrating 3D Terrain Information

[0065] 1. Model building

[0066] After constructing the 3D terrain fusion features, a spatial-spectral Mamba segmentation model incorporating 3D terrain information was built to achieve automatic identification and segmentation of pegmatite dikes. The Mamba model is a sequential modeling method based on a state-space modeling mechanism, capable of efficiently modeling long-distance dependencies with low computational complexity, and suitable for the fusion and representation of multi-dimensional feature information. Therefore, in large-scale remote sensing image segmentation tasks, the Mamba model can effectively extract the continuous distribution features of geological targets in both spatial and spectral dimensions.

[0067] In this invention, the constructed spatial-spectral Mamba network mainly includes a spectral feature extraction branch, a spatial feature extraction branch, and multiple spatial-spectral Mamba modules. The spectral branch is used to extract spectral feature information of multispectral images in the band dimension, the spatial branch is used to extract structural features of images in the spatial neighborhood, and the spatial-spectral Mamba modules jointly model long-distance dependencies in the spatial and spectral dimensions through a state-space modeling mechanism to achieve deep fusion of multidimensional features. Figure 5 A spatial-spectral Mamba model structure diagram that incorporates three-dimensional terrain.

[0068] Three-dimensional terrain information and multispectral image features are jointly input into a spatial-spectral Mamba model for feature encoding. Specifically, during the model input stage, elevation information from the digital elevation model is embedded as an independent channel into spatial feature tokens to characterize the geometric relationships and topographic continuity between different spatial locations. Simultaneously, spectral tokens are constructed from the information of each band of the multispectral image to describe the spectral variation characteristics between different bands at the same spatial location. Subsequently, the spatial tokens and spectral tokens are respectively input into a spatial-spectral Mamba encoder composed of multiple basic Mamba modules for feature extraction. Each basic Mamba module includes a linear mapping layer, a state-space model (SSM) unit, and a normalization structure. The SSM unit is used to model long-distance dependencies in the input sequence, thereby achieving efficient feature representation in both spatial and spectral dimensions. Through the stacking of multiple spatial-spectral Mamba modules, the spatial and spectral branches gradually achieve information interaction and fusion at the high-level feature stage, forming a joint feature representation that simultaneously contains spatial structure information and spectral discrimination information, providing a reliable feature foundation for subsequent pegmatite dike segmentation.

[0069] 2. Loss Function Setting

[0070] During model training, a loss function incorporating terrain information is introduced to constrain the network, further improving the consistency between the segmentation results and the real terrain spatial structure. The model's total loss function is defined as:

[0071] (4)

[0072] in, This is a supervised segmentation loss function based on labeled samples, used to constrain the difference between the model output and the true label; This is the elevation constraint loss term, used to incorporate terrain information to spatially constrain the segmentation results; The terrain weight parameter controls the degree of influence of terrain information during model training. This approach enables the model to fully utilize terrain elevation information while learning spectral and spatial features, thereby improving the accuracy and spatial continuity of pegmatite vein segmentation results under deeply dissected and complex terrain conditions.

[0073] 3. Dataset Construction and Model Training

[0074] Based on the results of field reconnaissance and lithological surveys of the study area, the main land cover types in the WorldView-3 multispectral imagery of the study area were interpreted and labeled, and a deep learning dataset for pegmatite vein segmentation was constructed. According to the geological characteristics and surface cover of the study area, the land cover was divided into four categories: Changchengian biotite interbedded with marble, Quaternary sediments, riverbed alluvial deposits, and pegmatite veins. Supervised annotation of image pixels was performed. Among these, pegmatite veins are mostly distributed in narrow strips, with a small spatial scale and are greatly affected by complex topography and shading, making them a target type with higher segmentation difficulty.

[0075] Regarding dataset partitioning, a spatial region-based dataset construction strategy was adopted, dividing the study area into several independent spatial regions. Approximately 80% of the region was used as the training set, and the remaining 20% ​​was used as the validation set. This approach avoids spatial overlap between training and validation samples, thereby improving the model's generalization ability. Detailed dataset information is shown in Table 1.

[0076] Table 1 Deep learning dataset for pegmatite vein segmentation experiment

[0077]

[0078] Subsequently, the spatial-spectral Mamba model incorporating 3D terrain information was trained using the constructed dataset. The model was trained for 200 epochs to achieve stable convergence of the model parameters. To verify the effectiveness of the proposed method, the model was compared with several typical classification and segmentation methods, including Support Vector Machine (SVM), K-Nearest Neighbors (KNN), 2D Convolutional Neural Network (2D CNN), 3D Convolutional Neural Network (3D CNN), Hybrid 2D-3D Convolutional Neural Network (HybridSN), and Spatial-Spectral Transformer (SSTN) models.

[0079] S4. Identification Result Analysis, Accuracy Evaluation and Field Verification

[0080] 1. Analysis of recognition results and evaluation of accuracy

[0081] In terms of model performance evaluation, overall accuracy (OA), average accuracy (AA), and Kappa coefficient are used as evaluation indicators to quantitatively evaluate the model segmentation results. The performance of the model in the pegmatite vein segmentation task is comprehensively evaluated from aspects such as overall classification consistency, class balance, and random consistency correction.

[0082] The comprehensive evaluation of recognition accuracy and segmentation results show that the spatial-spectral Mamba model fusion of 3D terrain information proposed in this application exhibits good performance across all accuracy metrics. For example... Figure 6 As shown, compared with traditional machine learning methods KNN and SVM, this method achieves significant improvements in overall accuracy (OA), average accuracy (AA), and Kappa coefficient, indicating that it can effectively improve the discrimination ability between different types of land features. Compared with typical deep learning models (2D-CNN, 3D-CNN, HybridSN, and SSTN), our methods maintain high classification accuracy while further enhancing the model's stability and recognition accuracy under complex terrain conditions.

[0083] In terms of classification accuracy, the method of this invention achieves an overall accuracy of 98.15%, an average accuracy of 98.30%, and a Kappa coefficient of 0.9725, all of which are superior to the comparative methods. Specifically, in the segmentation and identification of the target category pegmatite veins, the classification accuracy reaches 97.29%, indicating that this method can effectively improve the identification ability of narrow, elongated ore bodies. By introducing DEM topographic information into the Spatial-Spectral Mamba model framework, collaborative modeling of spatial, spectral, and topographic features is achieved, enabling the model to more accurately characterize the spatial differences between features under complex terrain conditions. Detailed accuracy figures are shown in Table 2.

[0084] Table 2 Comparison of Deep Learning Segmentation Accuracy

[0085]

[0086] from Figure 7 From the spatial distribution characteristics of the segmentation results, Quaternary sediments, Great Wall biotite marble, and alluvial deposits exhibit a relatively reasonable spatial distribution pattern within the study area, corresponding well to the actual geomorphological features of the region. The pegmatite dikes, in the segmentation results, generally show a continuous banded structure along a certain direction, consistent with the regional tectonic trend and mountain undulation. In areas with significant slope variations, the segmentation results still maintain good spatial continuity and boundary integrity, without obvious breaks or dispersion.

[0087] Figure 8 The results of two-dimensional and three-dimensional segmentation of pegmatite dikes in deeply incised areas are presented. Among all the comparison methods, the Spatial-Spectral Mamba model (i.e., SS-Mamba in Table 2) performs best in terms of classification accuracy; therefore, the two-dimensional segmentation results of this model are selected for comparison. The original Spatial-Spectral Mamba model does not incorporate a three-dimensional terrain embedding module and only performs feature learning based on spatial-spectral information. Figure 8 In the diagram, (a) represents the overall distribution of the two-dimensional segmentation results. Figure 8 Image (c) shows a magnified view of the two-dimensional segmentation results. It can be seen that without the introduction of three-dimensional topographic information, due to the dramatic topographic undulations and complex slope aspect variations in the study area, some dikes exhibit discontinuities or omissions in the segmentation results in areas 1, 3, and 5. The magnified view also shows that the two-dimensional results, relying solely on the spatial-spectral characteristics of the image, are insufficient to accurately reflect the true spatial morphology of the pegmatite dikes under complex topographic conditions. Figure 8 As shown in (b), after introducing a three-dimensional terrain model, the segmentation results are expanded from a two-dimensional plane to a three-dimensional space and can be mapped according to the terrain undulations, thus reflecting the actual distribution of the dikes under different elevations, slopes, and aspects. Compared with two-dimensional results, the three-dimensional representation can better maintain the spatial continuity of the dikes in areas with significant terrain changes such as steep slopes, valleys, and slope break zones, and reduce boundary offsets caused by terrain shading or projection deformation, while preserving the spatial connection between different elevation layers.

[0088] Figure 8 (d) shows a localized three-dimensional magnification of eight typical areas. The results indicate that the pegmatite dikes are generally distributed in banded, lenticular, or irregular strip patterns. Their extension direction is mostly consistent with the slope strike or regional tectonic direction, and they can exhibit bending, turning, or branching phenomena with topographic undulations. The three-dimensional segmentation results can more intuitively reflect the continuity of the dikes between different elevations and their dip changes on the slope, thus providing a more complete expression of the spatial distribution characteristics of the pegmatite dikes.

[0089] Therefore, the method of the present invention, by integrating the spatial-spectral information of hyperspectral imagery with DEM topographic information, can achieve high-precision identification and segmentation of pegmatite veins under complex terrain conditions. This not only improves the overall classification accuracy but also more realistically reflects the spatial distribution characteristics of the target ore body, thus providing an effective technical means for remote sensing identification and prospecting prediction of mineral resources in complex mountainous areas.

[0090] 2. On-site verification

[0091] To verify the practical application effect of the pegmatite vein segmentation method integrating terrain information described in this invention, the areas where two main pegmatite veins are located in the study area were selected ( Figure 8 Regions 1 and 2 in the diagram are used as the field verification area. The verification results are as follows: Figure 9 As shown, the comparison between remote sensing segmentation results and field observations reveals that the segmentation results accurately reflect the spatial location and extension direction of the dikes. Even in steep slopes and complex terrain areas, the dikes maintain their continuity along the slope, without large-scale shifts or systematic misclassifications. The local discontinuities or patchy phenomena are mainly due to the discontinuities and spectral differences within the dike outcrops themselves, rather than classification errors.

[0092] To further verify the mineral type, representative rock samples were collected and thin sections were prepared for microscopic analysis. The results showed that the samples were lithium-bearing pegmatite, with spodumene as the main mineral, consistent with known mineralization in the region, thus confirming the accuracy of the remote sensing segmentation. Combining the results of remote sensing segmentation and field verification, the method of this invention can accurately segment and identify the spatial distribution and extension trend of pegmatite veins under complex terrain conditions, demonstrating good stability and applicability.

[0093] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A remote sensing segmentation method for pegmatite veins in deeply dissected areas, incorporating three-dimensional terrain, characterized in that... The steps are as follows: S1. Acquire multispectral remote sensing images and lidar point cloud data of the study area, preprocess the multispectral remote sensing images, process the lidar point cloud data to generate a digital elevation model, and spatially register the digital elevation model with the preprocessed multispectral remote sensing images. S2. Construct a three-dimensional terrain information fusion module, introduce the elevation information in the digital elevation model as an independent dimension into the pixel feature space of the multispectral remote sensing image, and expand the original two-dimensional spatial-spectral features of each pixel into a three-dimensional feature representation that includes planar coordinates, elevation and spectral features. S3. Construct a spatial-spectral Mamba model that integrates three-dimensional terrain information. Input the three-dimensional feature representation obtained in step S2 into the spatial-spectral Mamba model to complete feature encoding and fusion. Use a total loss function that includes an elevation constraint term to supervise the training of the model. S4. Use the trained spatial-spectral Mamba model to segment and identify pegmatite veins in the study area, and output the remote sensing segmentation results of the pegmatite veins.

2. The remote sensing segmentation method for pegmatite veins in deeply cut areas based on three-dimensional terrain as described in claim 1, characterized in that: In step S2, the three-dimensional feature representation specifically includes: Let the spectral influence be defined in a two-dimensional spatial domain. The corresponding digital elevation model is a continuous elevation function. ,in, Indicates pixel position The surface elevation value at the location; each pixel consists of the original two-dimensional spatial-spectral features. Extended to three-dimensional feature representation , where s represents the spectral feature vector of the corresponding pixel.

3. The remote sensing segmentation method for pegmatite veins in deeply cut areas based on three-dimensional terrain as described in claim 1, characterized in that: In step S3, the space-spectral Mamba model includes a spectral feature extraction branch, a spatial feature extraction branch, and multiple cascaded space-spectral Mamba modules. In the model input stage, the elevation information of the digital elevation model is embedded as an independent channel into the spatial feature token, and the information of each band of the multispectral image is constructed into a spectral token, which is then input into the space-spectral Mamba encoder to complete feature extraction.

4. The remote sensing segmentation method for pegmatite veins in deeply cut areas based on three-dimensional terrain as described in claim 3, characterized in that: Each of the spatial-spectral Mamba modules includes a linear mapping layer, a state-space model unit, and a normalization structure. The state-space model unit is used to model long-distance dependencies of the input sequence, thereby achieving feature fusion of spatial and spectral dimensions.

5. The remote sensing segmentation method for pegmatite veins in deeply cut areas based on three-dimensional terrain as described in claim 1, characterized in that: In step S3, the total loss function is defined as: in, The supervised segmentation loss function is based on labeled samples. This is an elevation constraint loss term, used to incorporate terrain information to constrain the spatial continuity of the segmentation results. These are terrain weight parameters.

6. The remote sensing segmentation method for pegmatite veins in deeply cut areas based on fusion of three-dimensional terrain as described in claim 5, characterized in that: In step S3, the model training adopts a dataset construction strategy based on spatial region partitioning: the study area is divided into several independent spatial regions, one part of which is used as the training set and the remaining part is used as the validation set, with no spatial overlap between the training set and the validation set.

7. The remote sensing segmentation method for pegmatite veins in deeply cut areas based on three-dimensional terrain as described in claim 1, characterized in that: In step S1, the multispectral remote sensing image is a high-resolution multispectral satellite image, and the preprocessing includes radiometric calibration and atmospheric correction; the digital elevation model is generated from point cloud data acquired by UAV-borne or airborne lidar systems through trajectory calculation, filtering, and ground point extraction, and the digital elevation model is resampled to the same spatial resolution as the multispectral remote sensing image.

8. The remote sensing segmentation method for pegmatite veins in deeply cut areas based on three-dimensional terrain as described in claim 1, characterized in that: In step S4, after obtaining the remote sensing segmentation results of the pegmatite veins, the overall accuracy OA, average accuracy AA, and Kappa coefficient are used as evaluation indicators to quantitatively evaluate the accuracy of the model segmentation results. Field observations are conducted in the pegmatite vein area within the study area, and rock samples are collected for thin section microscopy analysis to complete the field verification of the segmentation results.

9. A remote sensing segmentation system for pegmatite veins in deeply dissected areas, incorporating three-dimensional terrain, for implementing the remote sensing segmentation method for pegmatite veins in deeply dissected areas, as described in any one of claims 1-8, characterized in that... include: The data acquisition and preprocessing module is used to acquire multispectral remote sensing images and lidar point cloud data of the study area, preprocess the multispectral remote sensing images, process the lidar point cloud data to generate a digital elevation model, and spatially register the digital elevation model with the preprocessed multispectral remote sensing images. The three-dimensional terrain fusion module is used to introduce the elevation information in the digital elevation model as an independent dimension into the pixel feature space of the multispectral remote sensing image, and to expand the original two-dimensional spatial-spectral features of each pixel into a three-dimensional feature representation that includes planar coordinates, elevation and spectral features. The model building and training module is used to build a spatial-spectral Mamba model that integrates three-dimensional terrain information. The three-dimensional feature representation is input into the spatial-spectral Mamba model to complete feature encoding and fusion. The model is trained under supervision using a total loss function that includes an elevation constraint term. The segmentation and recognition module is used to segment and recognize pegmatite veins in the study area using a trained spatial-spectral Mamba model, and output the remote sensing segmentation results of the pegmatite veins.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the remote sensing segmentation method for pegmatite veins in deeply cut areas that incorporates three-dimensional terrain, as described in any one of claims 1-9.

Images