A Machine Learning-Based Method for Lithology Identification of Outcrops Using TLS-Derived Spectral and Geometric Features

Abstract

Lithological identification of outcrops in complex geological settings plays a crucial role in hydrocarbon exploration and geological modeling. To address the limitations of traditional field surveys, such as low efficiency and high risk, we proposed an intelligent lithology recognition method, SG-RFGeo, for terrestrial laser scanning (TLS) outcrop point clouds, which integrates spectral and geometric features. The workflow involves several key steps. First, lithological recognition units are created through regular grid segmentation. From these units, spectral reflectance statistics (e.g., mean, standard deviation, kurtosis, and other related metrics), and geometric morphological features (e.g., surface variation rate, curvature, planarity, among others) are extracted. Next, a double-layer random forest model is employed for lithology identification. In the shallow layer, the Gini index is used to select relevant features for a coarse classification of vegetation, conglomerate, and mud–sandstone. The deep-layer module applies an optimized feature set to further classify thinly interbedded sandstone and mudstone. Geological prior knowledge, such as stratigraphic attitudes, is incorporated to spatially constrain and post-process the classification results, enhancing their geological plausibility. The method was tested on a TLS dataset from the Yueyawan outcrop of the Qingshuihe Formation, located on the southern margin of the Junggar Basin in China. Results demonstrate that the integration of spectral and geometric features significantly improves classification performance, with the Macro F1-score increasing from 0.65 (with single-feature input) to 0.82. Further, post-processing with stratigraphic constraints boosts the overall classification accuracy to 93%, outperforming SVM (59.2%), XGBoost (67.8%), and PointNet (75.3%). These findings demonstrate that integrating multi-source features and geological prior constraints effectively addresses the challenges of lithological identification in complex outcrops, providing a novel approach for high-precision geological modeling and exploration.

1. Introduction

Geological outcrops are surface exposures of stratigraphic units and ore bodies. Accurate lithological identification of outcrops is fundamental to geological mapping, stratigraphic–structural analysis, and the study of sedimentary reservoir development patterns [1]. Different lithologies exhibit significant variations in spatial distribution patterns and internal structural characteristics [2,3]. Traditional field investigations rely heavily on expert on-site interpretation, often supplemented by laboratory analyses such as thin-section petrography. However, these methods have limited capacity to capture the spatial heterogeneity of outcrop lithologies—especially in complex terrains like cliffs or landslide zones. In these environments, conventional approaches are not only inefficient but also pose significant safety risks, making them unsuitable for high-precision, high-efficiency lithological interpretation [4,5,6,7,8].

With advancements in remote sensing technologies, digital geological exploration has become a major research focus [9]. Terrestrial laser scanning (TLS), as a high-resolution 3D data acquisition technique, captures dense spatial geometry and spectral reflectance information, enabling rapid collection of 3D point cloud data from target areas [10]. Compared with wide-area observation techniques such as UAV photogrammetry (planimetric accuracy > 5 cm) [11] and satellite remote sensing (spatial resolution > 1 m) [9], TLS offers millimeter- to sub-millimeter-level accuracy [12], making it well-suited for the quantitative analysis of geological features [13]. As a result, TLS technology has been widely applied in geological monitoring [14], 3D geological modeling [15,16], and intelligent identification of rock mass structural planes [17,18], among other domains [19].

Although TLS-based 3D point clouds provide high data accuracy, the effective extraction of geological information—particularly for lithological interpretation—remains a significant challenge [20]. By leveraging differences in spectral reflectance among various minerals, multispectral LiDAR systems have been successfully employed in lithological and mineral identification tasks [21,22,23,24]. In addition, geometric descriptors have proven effective in characterizing rock layer textures, such as surface roughness and curvature [25], and have also been employed in lithology-related applications [26]. However, most existing approaches rely on a single type of feature—either spectral reflectance or geometric morphology descriptors—with relatively few studies investigating the integration of both. This limitation reduces the effectiveness of TLS point cloud applications in geologically complex environments [27]. As a result, the fusion of spectral and geometric features derived from TLS data has become a key research focus. For instance, Chen et al. [28] proposed a Spatial Case-Based Reasoning (SCBR) model that integrates historical stratigraphic spatial distribution patterns with physical attribute features to enable stratigraphic identification. Penasa et al. [29] demonstrated the effectiveness of multi-feature collaborative interpretation by identifying outcrop flint layers through a combination of TLS intensity texture and local geometric descriptors. Similarly, Weidner et al. [30] achieved high-accuracy classification of bedrock, gravel, vegetation, and snow cover on natural slopes under multi-seasonal conditions by integrating multi-scale geometric features (e.g., surface curvature, slope) with LiDAR echo intensity.

The rapid advancement of machine learning and deep learning techniques has significantly enhanced the accuracy and efficiency of automated lithological identification systems [22,29]. However, natural weathering processes and sensor-induced noise remain major factors limiting the accuracy of TLS-based lithology classification [31]. Ercoli et al. [32] found that the surface reflectance characteristics of limestone are significantly affected by weathering when measured using laser scanning. This suggests that data-driven models may produce classification results that deviate from geological reality, highlighting the necessity of incorporating geological prior knowledge to constrain the interpretation process. Buckley et al. [33] demonstrated that integrating field geological measurements (e.g., structural attitudes) can effectively improve systematic errors in LiDAR data and enhance the geological reliability of the resulting interpretations.

According to fundamental principles of stratigraphy, stratigraphic units in stable depositional environments should exhibit lateral continuity and conformable contacts [34]. However, these principles have rarely been incorporated into existing TLS-based lithological classification studies. Most current research has focused on algorithmic optimization to enhance classification robustness, such as continuous improvements in deep learning network architectures [1,35]. To address these limitations, this study proposes an improved lithological classification framework for terrestrial laser scanning (TLS) outcrop point clouds, hereinafter referred to as SG-RFGeo (Spectral–Geometric Random Forest with Geological Constraints). The framework includes the following steps:

(a)

Segmenting the outcrop point cloud into regular grid units;

(b)

Extracting a multidimensional feature set, including spectral features (e.g., reflectance median, mean, skewness) and geometric features (e.g., curvature, planarity, sphericity);

(c)

Constructing a random forest model based on the extracted features to perform lithological classification;

(d)

Applying post-classification corrections based on prior stratigraphic attitude information to enhance geological consistency.

Field validation conducted at the Yueyawan outcrop of the Qingshuihe Formation, located on the southern margin of the Junggar Basin, demonstrates that the SG-RFGeo method significantly outperforms traditional approaches in terms of lithological classification accuracy and stratigraphic consistency. These findings highlight the importance of integrating geological domain knowledge with machine learning techniques and offer a novel technical framework for lithology identification using TLS data. The main contributions of this research are as follows:

(a)

Grid-based multi-source feature fusion method: A systematic framework was developed for feature extraction from gridded point cloud units, enabling the simultaneous acquisition of spectral reflectance statistics (e.g., median, kurtosis) and 3D geometric descriptors (e.g., surface variation, curvature, planarity). This method supports a comprehensive, multidimensional characterization of outcrop point cloud attributes, providing a strong foundation for robust lithological classification.

(b)

Geology-constrained post-classification correction algorithm: A stratigraphic boundary-guided refinement strategy is proposed, incorporating geological prior knowledge—specifically, the stratigraphic principle of lateral continuity, which posits that lithologies within the same sedimentary layer tend to exhibit lateral consistency and internal homogeneity. By fitting stratigraphic surfaces and applying majority voting within each delineated layer, the method effectively reduces misclassifications caused by surface weathering or sensor-induced noise. This post-processing step significantly improves both the spatial continuity and geological plausibility of the classification results.

2. Data

2.1. Study Area

The study area is located at Yueyawan on the southern margin of the Junggar Basin (Figure 1b), Xinjiang, where the exposed strata belong to the Lower Cretaceous (LC) Qingshuihe Formation (K1q). This formation predominantly consists of three lithologies: mudstone, sandstone, and conglomerate. The lower part of the outcrop consists of a sequence of fan delta deposits, including distributary channel conglomerates of the fan delta plain (DP) (Figure 1c(②)), gravelly sandstones of underwater distributary channels at the fan delta front (FDF), and sandbar sandstones of the fan delta mouth bar (FDMB) (Figure 1c(①)). The middle to upper sections of the outcrop are dominated by shallow lacustrine deposits, characterized by thinly interbedded sandstone and mudstone layers (Figure 1c(③,④)).

Overall, the outcrop in the study area is well-exposed, featuring distinct lithological assemblages and clear stratigraphic contacts, which provide favorable regional conditions for lithological identification. However, the exposed strata exhibit highly variable depositional environments and complex lithofacies associations. The outcrop includes a diverse range of lithologies with minimal color contrast—primarily grayish-white—which pose significant challenges for conventional automated lithological recognition. Furthermore, the rocks exhibit low degrees of diagenesis and have undergone prolonged exposure, resulting in significant weathering and surface loose materials. These factors further complicate accurate lithological classification.

2.2. Data Acquisition

In this study, a RIEGL VZ-400 terrestrial laser scanning system (RIEGL Laser Measurement Systems GmbH, Horn, Austria) was used, with its key technical specifications summarized in

Table 1. Technical specifications of the laser scanning instrument.

Model	Pulse Frequency (kHz)	Data Acquisition Rate (pts/s)	Scanning Speed (lines/s)	Field of View (°)	Range Accuracy (mm @ m)	Angular Resolution (°)
RIEGL VZ400	1200	500,000	<100	360 × 100	±5 mm @ 50 m	<0.001

. This instrument operates using a pulsed laser beam in the near-infrared (NIR) spectrum, offering high sensitivity to NIR-responsive surface materials. It is suitable for a wide range of field conditions, providing a maximum measurement range of up to 1000 m under favorable weather and unobstructed line-of-sight conditions. The system is equipped with a dual-processing platform that enables real-time data acquisition and on-site registration, eliminating the need for ground control targets. This capability significantly enhances both data acquisition efficiency and 3D reconstruction accuracy.

The experiment was conducted using a reference scanning distance of 10 m with a planar measurement accuracy of ±1 mm. To accommodate the scanner’s built-in tilt compensation limit (±10°), scanning stations were strategically positioned at elevated locations within the study area to ensure an unobstructed line of sight to the outcrop face. The data acquisition workflow began with an initial 360° panoramic pre-scan of the target outcrop, which was then used to guide high-precision directional scans. Simultaneously, high-resolution frontal digital images of the outcrop were captured from multiple viewpoints, ensuring a minimum overlap of 15% between adjacent images. This overlap was maintained to support subsequent texture mapping of the point cloud and to facilitate geological interpretation.

The outcrop point cloud data were processed using RiSCAN PRO (v2.11), proprietary software developed by RIEGL that integrates both data pre-processing and advanced analytical functionalities. Within this platform, multi-station registration and outlier filtering were performed, resulting in a final dataset comprising 56,063,397 valid point samples.

In this study, a total of 174 standard rock samples were systematically collected from the study area for laboratory-based thin-section analysis. Lithological identification classified the samples into four categories: conglomerate, sandstone, mudstone, and vegetation, covering all lithological types present within the study area. A summary of the manually labeled point cloud samples corresponding to these categories is presented in

Table 2. The number of samples of the collected dataset.

Class	Samples
Vegetation	3369
Conglomerate	3800
Sandstone	5500
Mudstone	4899

.

As shown in Table 2, a total of 17,568 point cloud samples were manually annotated in this study. Among the lithological classes, sandstone accounted for the most significant proportion (31.3%), followed by mudstone (27.9%) and conglomerate (21.6%), which is consistent with the actual lithological distribution observed in the field. Vegetation, which is not the primary focus of this study and differs significantly from rock types in its characteristics, accounted for the smallest proportion at 19.2%. The overall sample distribution is relatively balanced, with a maximum-to-minimum class ratio of 1.63:1, effectively reducing the risks of model bias and overfitting.

3. Methods

To address the challenges of lithology recognition in outcrop point clouds under complex geological conditions, this study proposes a classification framework named SG-RFGeo, which integrates spectral–geometric hybrid features. The method consists of four main modules: (a) regular grid segmentation, (b) spectral and geometric feature extraction, (c) double-layer random forest classification, and (d) post-processing based on geological constraints. The overall technical workflow is illustrated in Figure 2.

Based on the pre-processed outcrop point cloud data, the SG-RFGeo method begins by dividing the study area into regular grid units, each with a fixed size of 10 mm × 10 mm × 10 mm. Within each unit, two categories of features were systematically extracted: 17 geometric descriptors and 8 spectral statistical features. An optimal subset of these features, selected using feature selection techniques, was then input into a hierarchical (double-layer) random forest classification model. The shallow layer performed a coarse classification to distinguish conglomerate, sand–mudstone, and vegetation. The deep layer refined this by further separating sandstone from mudstone within the sand–mudstone class. Finally, a post-processing step, guided by stratigraphic consistency, was applied to refine the classification results. This ensured lithological continuity within individual stratigraphic layers and enhanced the overall accuracy of the outcrop lithology identification.

3.1. Regular Grid Partitioning of Point Cloud Data

The construction of lithological identification units provides the foundation for extracting grid-based feature values. In this study, a regular grid partitioning strategy was adapted—a widely used method in 3D point cloud segmentation. This approach divides the spatial domain of the point cloud into uniformly sized cubic cells, with each assigned a unique spatial identifier. Specifically, the spatial extent of the outcrop point cloud is first delineated, and a bounding cube is defined to represent the practical analysis region. Based on a predefined grid resolution (i.e., the side length of each cube), the region is subdivided into a three-dimensional grid structure. All points falling within the same grid cell are grouped into a single lithological identification unit, which serves as the fundamental element for subsequent feature extraction and classification. The spatial partitioning scheme is illustrated in Figure 3.

The grid cell size is a critical parameter in grid-based segmentation, as it directly affects both the feature representativeness of extracted features and the spatial resolution of the dataset. In theory, a larger grid cell contains more points, which improves the statistical stability of feature extraction and enhances the differentiation of laser reflectance intensities that arise from variations in the petro-physical properties of different rock types. However, a huge cell size can significantly reduce the spatial resolution of the dataset, potentially leading to the loss of crucial local feature details.

Lithological properties are generally consistent within a single stratigraphic unit but can vary significantly between different units. If the grid size exceeds the thickness of thinly stratified layers, a “cross-layer” effect may occur, wherein a single cell contains multiple lithologies. This leads to signal mixing, which compromises the ability of both reflectance and geometric features to accurately represent the true lithological characteristics of the outcrop. Therefore, selecting an appropriate grid size requires careful consideration of both geological conditions and instrument capabilities. Ideally, the grid resolution should correspond to, or be smaller than, the minimum thickness of the exposed stratigraphic layers.

At the Yueyawan outcrop, visual interpretation revealed that the thinnest mud–sand interbeds are approximately 1 cm thick. Considering the scanner’s step size and data density, a final grid resolution of 10 × 10 × 10 mm was selected. This configuration achieves a balance between spatial resolution and feature robustness, providing a reliable foundation for subsequent feature extraction and lithological classification.

3.2. Geometric–Spectral Feature Extraction

Accurate feature representation plays a pivotal role in determining the performance of lithological classification. The effective extraction of lithology-sensitive features is a critical prerequisite for building a robust classification model. In this study, the outcrop point cloud data contained both three-dimensional spatial coordinates and laser reflectance intensity values, providing a rich source of geometric and spectral information for feature derivation.

Accordingly, two categories of feature sets were systematically developed: geometric morphological features and spectral statistical features. The geometric features capture the spatial structure and surface complexity of the outcrop, while the spectral features reflect variations in laser reflectance associated with mineral composition and surface properties. By adopting a multidimensional feature fusion strategy, this study integrates complementary information from both spatial and spectral domains, aiming to identify feature combinations that are highly sensitive to lithological variation. This approach provides a solid foundation for subsequent machine learning-based classification.

3.2.1. Geometric Features

The surface geometric morphology of rocks is closely related to both internal and external factors, including mineral grain size, spatial arrangement patterns, diagenetic pressure, temperature, and fluid activity. The interaction of these factors produces significant variations in surface geometry across rocks of different lithologies [36]. Taking clastic rocks as an example, conglomerates are typically formed in high-energy depositional environments and are composed primarily of coarse-grained mineral debris. Hydraulic sorting leads to the development of well-defined intergranular pores among the coarse clasts, resulting in irregular surface geometries characterized by high roughness, low planarity, and strong anisotropy. In contrast, mudstones are deposited in low-energy, still-water environments and are composed of fine-grained argillaceous materials. Prolonged and intense compaction transforms these sediments into dense, homogeneous structures with relatively smooth surfaces. As a result, the geometric features are characterized by low roughness and high planarity.

Based on the high-resolution 3D point cloud data, a total of 17 geometric feature parameters were derived through grid-based analysis [37,38]. These parameters include mean curvature, Gaussian curvature, normal change rate, eigenvalue-based descriptors (eigenvalue1, eigenvalue2, eigenvalue3), principal component parameters (PCA1 and PCA2), sum of eigenvalues, omnivariance, eigenentropy, anisotropy, planarity, linearity, surface variation, sphericity, and verticality. The mathematical definitions and representative visualizations of each parameter are provided in

Table 3. Geometric feature parameters. The figure presents the aforementioned 14 geometric feature parameters of point cloud data. Each parameter is accompanied by its mathematical formula and a corresponding visualization derived from typical outcrop point cloud data, which assists in lithological classification and structural analysis. The visualizations are rendered using pseudocolor, where red indicates high feature values, blue indicates low values, and yellow-green tones represent intermediate values.

	Feature Parameters	Formulas	Visualizations		Feature Parameters	Formulas	Visualizations
1	Mean curvature	$H={\displaystyle \frac{k_1+k_2}{2}}$		2	Gaussian curvature	$K_g=k_1\times k_2$
3	PCA (PCA1, PCA2)	$\mathrm{P}\mathrm{C}\mathrm{A}=u^T·C·u$		4	Sum of eigenvalues	$S_{\lambda }={\lambda }_1+{\lambda }_2+{\lambda }_3$
5	Anisotropy	$A={\displaystyle \frac{{\lambda }_1-{\lambda }_3}{{\lambda }_1}}$		6	Planarity	$P={\displaystyle \frac{{\lambda }_2-{\lambda }_3}{{\lambda }_1}}$
7	Linearity	$L={\displaystyle \frac{{\lambda }_1-{\lambda }_2}{{\lambda }_1}}$		8	Sphericity	$S_{ph}={\displaystyle \frac{{\lambda }_3}{{\lambda }_1}}$
	Feature Parameters		Formulas				Visualizations
9	Verticality		$V=1-{\left|n\times z\right|}^c$
10	Normal change rate		$NCR={\displaystyle \frac{1}{N}}{\sum }_{i=1}^N(1-|n_p\times n_i|)$
11	Eigenvalue (λ1, λ2, λ3)		$C={\displaystyle \frac{1}{n}}\sum \left(X_i-\mu \right){\left(X_i-\mu \right)}^T$ $C·u=\lambda ·u$
12	Omnivariance		$\mathrm{O}=\sqrt[3]{{(\lambda }_1\times {\lambda }_2\times {\lambda }_3)}$
13	Eigenentropy		$\mathrm{E}=-{\sum }_{i=1}^3{\lambda }_i\ln {\lambda }_i$
14	Surface variation		$SV={\displaystyle \frac{{\lambda }_3}{{\lambda }_1+{\lambda }_2+{\lambda }_3}}$

Note: Although 14 rows are shown, some entries include multiple sub-features. Specifically, the PCA feature contains two components (PCA1and PCA2), and the eigenvalue feature includes three components (λ₁, λ₂, and λ₃). Their visualizations are stacked within a single cell, from bottom to top: PCA1 followed by PCA2, and λ₁ followed by λ₂ and λ₃, respectively. In total, 17 geometric features were used in the analysis.

.

The aforementioned multidimensional features effectively capture the geometric properties of rock surfaces. Beyond representing the spatial distribution of grain sizes, these features also implicitly encode critical geological information, including diagenetic environments, depositional processes, and mineralogical composition. As a result, they provide a robust quantitative foundation for distinguishing between clastic lithologies (e.g., conglomerates and mudstones), thereby significantly improving the accuracy and reliability of automated lithological classification.

3.2.2. Spectral Features

Reflectance describes a surface’s ability to reflect electromagnetic waves at various wavelengths and is influenced by a combination of factors, including chemical composition, physical condition, and surface morphology [39]. Franceschi et al. [40] reported a strong negative correlation (r = −0.85) between reflectance and clay content in rocks when illuminated by a laser at 1535 nm. Furthermore, the reflectance values of limestone, marl, and clay were found to differ significantly from those measured under natural field conditions, underscoring the potential of laser-derived reflectance as a reliable discriminative feature for lithological classification.

Compared to geometric features that characterize the macroscopic structure of outcrops, spectral reflectance provides complementary discriminative information at the microscopic scale by capturing compositional variations among rock types. Many common minerals exhibit diagnostic absorption features within the visible to near-infrared (VNIR) range (400–2500 nm), such as Fe²⁺ charge transfer absorption near 900 nm and OH vibrational absorption of clay minerals around 2200 nm [41]. In regions with complex mineralogical compositions and well-developed weathering crusts, spectral reflectance offers a significant advantage by improving the accuracy of lithological classification.

In this study, multiple reflectance-based statistical features were extracted from the segmented grid cells to construct a comprehensive reflectance feature set. This set includes eight parameters: maximum, minimum, mean, median, standard deviation, skewness, kurtosis, and coefficient of variation. The corresponding formulas and local visual representations of each parameter are presented in

Table 4. Reflectivity statistical parameters. The figure presents the aforementioned 8 spectral statistical feature parameters of point cloud data. Each parameter is shown with its mathematical formula and a corresponding visualization derived from typical outcrop point cloud data, which assists in lithological classification and structural analysis. The visualizations are rendered using pseudocolor, where red indicates high feature values, blue indicates low values, and yellow-green tones represent intermediate values.

	Feature Parameters	Formulas	Visualizations		Feature Parameters	Formulas	Visualizations
1	Max	$Max=\max \left(I_i\right)$		2	Min	$Min=\min \left(I_i\right)$
3	Mean	$\overline{x}={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{I}_i$		4	Coefficient of Variation	$c_v={\displaystyle \frac{\overline{i}}{\sigma }}$
	Feature Parameters	Formulas				Visualizations
5	Median	$M=\left\{\begin{array}{cc}{I}_{{\displaystyle \frac{\left(N+1\right)}{2}}},\hfill & N is odd\\ {[I}_{\left({\displaystyle \frac{N}{2}}\right)}+I_{\left({\displaystyle \frac{N}{2}}+1\right)}]{\displaystyle \frac{1}{2}},& N is even\end{array}\right.$
6	Standard Deviation	$\sigma =\sqrt{{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{(I_i-\overline{i})}^2}$
7	Skewness	$C_s={\displaystyle \frac{\frac{1}{N}{\sum }_{i=1}^N{(I_i-\overline{i})}^3}{{\sigma }^3}}$
8	Kurtosis	$K={\displaystyle \frac{\frac{1}{N}{\sum }_{i=1}^N{(I_i-\overline{i})}^4}{{\sigma }^4}}-3$

.

Extreme value parameters (maximum and minimum) are effective for distinguishing lithologies with substantial differences in reflectance. The mean and median represent the central tendency of the reflectance distribution, reflecting the overall reflectance properties of the rock surface. Standard deviation measures the degree of reflectance variability. Skewness captures the asymmetry of the distribution, while kurtosis describes the degree of peakedness: higher kurtosis values indicate highly concentrated reflectance distributions (homogeneous materials), whereas lower values reflect more dispersed distributions (heterogeneous structures). By integrating these statistical features, a more comprehensive and robust interpretation of rock surface reflectance can be achieved.

3.3. Random Forest-Based Lithological Classification

Given that the grid-based features integrate multimodal geometric and spectral statistical information, a random forest (RF) algorithm was employed to construct the lithological classification model. As an ensemble learning method composed of multiple decision trees, the random forest offers strong classification performance and robustness against overfitting, making it particularly well-suited for handling high-dimensional feature spaces. To further improve the classification accuracy, a two-stage classification framework was developed. In the first stage, highly discriminative features were selected using the Gini index to perform coarse classification among vegetation, conglomerate, and mud–sandstone units. In the second stage, an optimized feature set was derived based on the results of the initial classification, enabling fine-grained discrimination between sandstone and mudstone —two classes that are often confused due to their similar properties.

3.3.1. Feature Selection

The optimized selection of input features plays a critical role in determining the accuracy and generalization performance of the random forest-based lithological classification model. In this study, key differences in spectral and geometric characteristics among outcrop lithologies were systematically investigated, with particular emphasis on reflectance and surface morphological features. To identify the most impactful variables, the built-in feature importance evaluation method of the random forest algorithm was employed. During model training, the random forest model evaluates the contribution of each feature to lithological identification based on the Gini impurity, which is mathematically expressed as

$$Gini\left(D\right)=1-{\sum }_{i=1}^C$$

(1)

where D denotes the dataset, C is the total number of classes, and P_i represents the proportion of samples belonging to class i within the dataset. The Gini impurity measures the degree of class impurity at a given node, and the reduction in this impurity resulting from a feature-based split serves as an indicator of that feature’s importance. A lower Gini impurity indicates a purer node, suggesting that the corresponding feature plays a more important role in classification.

To effectively identify and retain the most informative features based on their Gini-derived importance scores, a Cumulative Importance Thresholding method was employed. In this approach, features are first ranked in descending order according to their importance. Their cumulative contributions to classification performance are then calculated sequentially. Once the cumulative importance exceeds a predefined threshold (typically set at 90%), the corresponding subset is selected as the dominant feature set. This strategy aims to capture the majority of decision-relevant information while minimizing feature redundancy. It enhances the generalization ability and robustness of the lithological classification model, while also maintaining an optimal balance between computational efficiency and accuracy.

3.3.2. Double-Layer Random Forest Model Construction

Building on the optimized feature selection and an importance ranking strategy, SG-RFGeo constructs a geometric–spectral feature framework tailored for complex outcrop environments. Given the significant differences in feature dimensions between sedimentary rocks and vegetation, a single-layer random forest often fails to effectively capture complex feature relationships. To enhance the model’s performance in complex geological settings, a two-stage classification architecture is proposed.

Each stage of the model employs an ensemble of decision trees, with lithological predictions determined through a majority voting mechanism. In the first stage, a random forest classifier is trained using core features selected based on the Gini index to perform coarse lithological classification. The target classes include vegetation, conglomerate, and mud–sandstone (encompassing both mudstone and sandstone), and were chosen to simplify the feature space and reduce model complexity. The second stage focuses on distinguishing sandstone from mudstone, particularly under the influence of thin interbedding effects. An optimized, high-sensitivity feature set and reconfigured model parameters were adopted to enhance discrimination accuracy for these commonly confused lithologies.

This hierarchical processing strategy facilitates progressive feature filtering and iterative model optimization through structured feature handling and adaptive parameter tuning. As a result, the model’s capacity to represent and interpret complex feature spaces is significantly enhanced.

To effectively model the nonlinear relationships between spectral and geometric features, a hyperparameter optimization strategy was employed using grid search in conjunction with five-fold cross-validation. This process aimed to fine-tune key random forest parameters, including the number of decision trees (n_estimators), maximum tree depth (max_depth), minimum number of samples per leaf node (min_samples_leaf), and the feature-splitting threshold. This systematic tuning ensures that the model achieves an optimal balance between accuracy, generalization, and computational efficiency.

Given the varying complexity of lithological classification tasks across different hierarchical levels, a layered parameter optimization approach was adopted to tailor the model configuration to each stage’s specific classification requirements.

Table 5. Hyperparameter settings for the double-layer random forest lithology classification model. This table presents the hyperparameter settings of the double-layer random forest lithology classification model, including parameter configurations for both the coarse classification (shallow-layer) and fine classification (deep-layer) stages. Key parameters such as the number of trees (n_estimators), maximum depth (max_depth), and minimum samples per leaf node (min_samples_leaf) are listed along with their values. The purposes of these settings are also explained, aiming to balance model accuracy and computational cost, prevent overfitting, and enhance the model’s sensitivity to subtle class differences.

Stage	Parameter	Value	Purpose
Coarse classification (shallow layer)	n_estimators	100	Balance accuracy and computation cost
	max_depth	20	Prevent overfitting while capturing key features
	min_samples_leaf	15	Ensure statistical robustness at each leaf node
Fine classification (deep layer)	n_estimators	200	Improve sensitivity to subtle class differences
	max_depth	30	Enhance feature representation capability
	min_samples_leaf	15	Maintain consistency of decision rules

summarizes the hyperparameter settings for the double-layer random forest classifier. In the shallow layer, parameters were chosen to enable efficient coarse discrimination, while in the deep layer, the settings were adjusted to enhance sensitivity for the more challenging fine-grained lithological classification. The rationale for each parameter choice is provided in the corresponding table column to support reproducibility and clarity.

In both stages of the classification framework, decision tree construction was guided by the Gini impurity criterion. Optimal feature-splitting thresholds were identified by maximizing node purity, thereby improving the alignment between predicted and actual lithological distributions within the feature space.

This double-layer modeling strategy significantly improves the lithological classification performance of the random forest model by leveraging multi-feature integration. It also effectively reduces overfitting risk by employing task-specific hyperparameters. Furthermore, the implementation of majority voting at the leaf node level enhances decision consistency and strengthens the model’s ability to accurately classify lithologies, especially in geologically complex regions where spectral and geometric characteristics overlap within outcrop point cloud data.

3.4. Post-Processing of Stratigraphic Attitude Results

Although the double-layer random forest model captures the general pattern of lithological distribution, its classification accuracy remains limited due to disturbances such as vegetation cover, weathering, and surface soil accumulation. To address these challenges, stratigraphic development parameters such as layer thickness and attitude are incorporated for post-classification optimization steps. The approach aims to improve the spatial continuity of lithological identification and ensure conformity with geological principles. According to stratigraphic theory, sedimentary layers within the same outcrop typically exhibit nearly parallel spatial orientations, and lithological properties within a single stratum tend to be internally homogeneous. Leveraging these geological constraints, this study introduces a probabilistic post-processing algorithm to refine the lithological classification results in outcrop point clouds by enforcing spatial consistency. A conceptual schematic of the proposed method is shown in Figure 4.

Figure 4 illustrates the abstraction of the actual outcrop surface into a mathematical surface model. The point cloud is partitioned into distinct classification units through regular grid segmentation. Each unit is classified by the two-tier random forest model and assigned a unique lithological label, represented by different colors in the figure. Stratigraphic boundary planes are fitted to group units belonging to the same stratigraphic layer. Based on the stratigraphic principle of lithological homogeneity within individual layers, the classification results are subsequently refined to enhance geological accuracy and consistency. The detailed procedure is illustrated in the flowchart below.

Figure 5 illustrates the complete workflow of lithology aggregation post-processing. The entire outcrop point cloud serves as the initial input. Based on the observed stratigraphic development characteristics, geologists manually select representative point clouds along stratigraphic boundaries using point cloud editing software (e.g., CloudCompare v2.13), and fit stratigraphic planes accordingly. For each fitted plane, the spatial relationship between the plane and the grid cells is analyzed to identify the set of cells intersected by the plane, which are then grouped as a single stratigraphic unit. The initial lithological classification results of all grid cells within this unit are then statistically analyzed, and the dominant lithology is determined by majority voting. This dominant category is reassigned to all grid cells within the group. This process is iteratively applied across the entire outcrop along the normal direction of the stratigraphic planes to achieve spatial aggregation and consistency correction of the lithological classification results. The implementation codes supporting this workflow are provided in the Supplementary Materials. The following outlines several key steps of the process:

Step 1: Fitting a Stratigraphic Plane Based on 3D Interface Data

To account for the 3D spatial distribution of outcrop point clouds, a mathematical model of the stratigraphic interface is first established. A representative stratigraphic unit is selected to characterize the local bedding orientation, and a best-fit plane is generated using the least squares method.

Let the selected point cloud contain N points, with spatial coordinates (x_i, y_i, z_i), i = 1, 2, …, N, to fit a plane of the following form:

Z = Ax + By + D

(2)

This problem can be expressed in matrix form as follows:

$$\mathrm{A}=\left[\begin{array}{ccc}{x}_1& y_1& 1\\ x_2& y_2& 1\\ ⋮& ⋮& ⋮\\ x_N& y_N& 1\end{array}\right],\mathrm{b}=\left[\begin{array}{c}{z}_1\\ z_2\\ ⋮\\ z_{\mathrm{N}}\end{array}\right]{\Rightarrow \left(A B D\right)}^{\mathrm{T}}={\left(A^{\mathrm{T}}A\right)}^{-1}{A}^{\mathrm{T}}\mathrm{b}$$

(3)

Solving this system using least squares yields the optimal coefficients A, B, and D, for the fitted plane, which serves as a geometric reference for subsequent lithological correction. In stratigraphic regions exhibiting pinch-outs or discontinuities, a piecewise fitting strategy or a locally weighted regression approach is employed to effectively address spatial irregularities and ensure accurate modeling of complex stratigraphic surfaces.

Step 2: Extraction of Iso-stratigraphic Point Cloud Data

Based on the fitted plane parameters A, B, D, the relative spatial position of each grid cell’s point set P with respect to the plane is determined using the following function:

$$\mathrm{f}\left(x_i,y_i,z_i\right)=\left\{\begin{array}{cc}A{x}_i+B{y}_i-z_i+D>0& \hfill \mathrm{A}\mathrm{b}\mathrm{o}\mathrm{v}\mathrm{e} \mathrm{t}\mathrm{h}\mathrm{e} \mathrm{p}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{e}\\ A{x}_i+B{y}_i-z_i+D=0& \hfill \mathrm{O}\mathrm{n} \mathrm{t}\mathrm{h}\mathrm{e} \mathrm{p}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{e}\\ A{x}_i+B{y}_i-z_i+D<0& \hfill \mathrm{B}\mathrm{e}\mathrm{l}\mathrm{o}\mathrm{w} \mathrm{t}\mathrm{h}\mathrm{e} \mathrm{p}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{e}\end{array}\right.$$

(4)

A grid cell is considered to intersect the plane (i.e., belong to a specific stratigraphic layer) if there exist at least two points $\exists p,q\in P$, s.t. $\mathrm{f}\left(p\right)·\mathrm{f}\left(q\right) \le 0$. This condition indicates that the cell contains points located on both sides of the plane, suggesting it intersects the stratigraphic boundary. To achieve complete traversal of the outcrop point cloud, the fitted plane is translated along its normal vector $\overrightarrow{n}=\left(A,B,-1\right)$ in steps equal to the grid resolution. This grid-based sweeping strategy facilitates the systematic extraction of point cloud subsets corresponding to iso-stratigraphic surfaces throughout the entire outcrop.

Step 3: Lithology Determination and Correction within Stratigraphic Layers

A majority voting mechanism is applied to the classification results of individual grid cells within each stratigraphic unit. The lithology category with the highest frequency in the unit is adopted as the final classification result, while the remaining categories are considered misclassifications and corrected accordingly. Specifically, for a given stratigraphic layer that intersects N grid cells, let the set of predicted lithology labels be {l₁, l₂, l₃, l₄} with corresponding frequencies {f₁, f₂, f₃, f₄}. The dominant lithology l^∗ selected by the majority voting method is then determined as follows:

$$l^{\ast }=arg \underset{l_i}{\max } f_i$$

(5)

To avoid misinterpretation of stratigraphic units, vegetation is intentionally excluded from being assigned as the final category during post-processing. If vegetation were selected as the dominant class within a stratigraphic layer, the entire layer would be classified as vegetation, which does not reflect geological reality. Since vegetation merely covers the outcrop surface rather than representing a lithological unit, such a result would obscure the true distribution of subsurface rock types and hinder geological interpretation. Therefore, when vegetation appears as the most frequent class within a layer, the lithology with the second-highest frequency is selected instead.

4. Results and Analysis

4.1. Evaluation Metrics

To systematically evaluate the performance of the lithology identification model based on outcrop point cloud data, a multidimensional assessment framework was established using the confusion matrix. The confusion matrix is an $n\times n$ matrix structure (where n denotes the number of classes), as illustrated in

Table 6. Confusion matrix.

	Predicted Positive	Predicted Negative
Actual Positive	TP	FN
Actual Negative	FP	TN

. The rows represent the true classes, while the columns correspond to the predicted classes. By comparing the predicted labels with the ground truth, the classification performance of the model can be quantitatively assessed.

In the table above, TN (true negative) refers to the number of negative samples correctly identified by the model, FP (false positive) represents the number of negative samples incorrectly classified as positive by the model, FN (false negative) indicates the number of positive samples that were missed by the model, and TP (true positive) denotes the number of positive samples correctly identified by the model.

Based on the confusion matrix, the following evaluation metrics are primarily adopted in this study:

(a)

Overall Accuracy (OA)

Overall accuracy is one of the most fundamental metrics for evaluating classification models. It represents the proportion of correctly classified samples to the total number of samples, providing an intuitive measure of the model’s overall classification performance. The calculation formula is

$$OA={\displaystyle \frac{{\sum }_{i=1}^k{TP}_i}{N}}$$

(6)

where TP_i (true positive) denotes the number of samples correctly classified as class i, k is the total number of classes, and N is the total number of samples in the test set. The value of OA ranges from 0 to 1, with a higher value indicating better overall classification performance.

(b)

Class-wise Evaluation Metrics

i

Precision measures the proportion of true positive instances among all instances predicted as positive.

$${Precision}_i={\displaystyle \frac{{TP}_i}{{TP}_i+{FP}_i}}$$

(7)

ii

Recall measures the proportion of true positive instances among all actual positive instances.

$${Recall}_i={\displaystyle \frac{{TP}_i}{{TP}_i+{FN}_i}}$$

(8)

iii

The F1-score is the harmonic mean of precision and recall used to address class imbalance issues.

$${F1}_i={\displaystyle \frac{2{Prec}_i\ast {Recall}_i}{{{Prec}_i+Recall}_i}}$$

(9)

To avoid dominant classes from skewing the evaluation results, this study uniformly adopts macro-averaged metrics to assess the balance of classification performance across different lithology types. The F1-score is selected as the core evaluation metric, supplemented by the Macro F1-score to evaluate the model’s recognition performance for each lithology. By combining precision and recall through the harmonic mean, the F1-score ensures classification accuracy while effectively controlling the risk of omission. This makes it particularly suitable for geological classification scenarios characterized by class imbalance and high feature correlation.

In this study, five-fold cross-validation was employed to evaluate the performance of the lithological classification model. This method effectively reduces the random errors introduced by a single data split, thereby improving the stability of evaluation metrics and the generalization ability of the model. In each iteration of the cross-validation, the dataset was randomly divided into a training set (80%) and a testing set (20%), and the evaluation metrics were computed accordingly. The final model performance was assessed by averaging the results of the five validation rounds.

4.2. Feature Importance Analysis

To further evaluate the contributions of different input features to the model’s classification performance, we analyzed the importance of the extracted spectral and geometric features using the Gini importance scores derived from the random forest model (as described in Section 3.3.1). This analysis helps identify which features are most discriminative across the four lithological classes and provides insights into their geological significance. Figure 6 presents the feature importance rankings for all 25 features, including spectral statistics and geometric descriptors, as determined by the random forest model.

As shown in Figure 6 below, overall, the spectral reflectance statistical features ranked higher in terms of their contribution, while geometric shape features also demonstrated stable contributions across multiple dimensions. The two feature types exhibit strong complementarity in the classification process. Specifically, the minimum reflectance ranked highest in importance, indicating significant differences in the minimum reflectance values between different targets within local regions. The standard deviations of reflectance, planarity, and median reflectance formed the second gradient of feature importance, representing data dispersion, local surface geometry, and the central trend of intensity distribution, respectively. The geological significance of these top-ranked features will be discussed in detail in Section 5, in conjunction with the physicochemical properties of the rocks and their depositional settings. In comparison, features such as Gaussian curvature, kurtosis, and verticality showed lower contributions to lithological identification in the study area. This ranking provides a crucial basis for feature selection and model optimization.

To further evaluate the discriminative power and class sensitivity of the most important features, the top four—minimum reflectance, standard deviation, median reflectance, and planarity—were selected for visualization analysis. Violin plots illustrating the distribution of these features across different classes are presented in Figure 7.

Figure 7 illustrates the distributions of the minimum reflectance, standard deviation, median reflectance, and planarity features across gravel, sandstone, mudstone, and vegetation samples. Notably, distinct differences are observed between vegetation and sedimentary rock classes in all four feature distributions. The violin plots for vegetation exhibit broad widths and a wide range of extreme values—particularly for minimum reflectance and standard deviation—characterized by noticeable long tails, which reflect the high dispersion of these features.

The planarity feature is highly concentrated in the low-value range, with a narrow and unimodal distribution, indicating the irregular geometric structure of vegetation point clouds. Within the sedimentary rock classes, gravel exhibits lower minimum reflectance and higher standard deviation compared to mudstone and sandstone, indicating greater geometric variability and spectral reflectance diversity, which facilitates its differentiation.

In contrast, the distributions of sandstone and mudstone are more similar, with smoother curves for minimum reflectance and standard deviation, indicating more concentrated feature values. The violin plot of mudstone displays a symmetrical, bell-shaped pattern centered around the median, suggesting that the median reflectance is stable and approximately normally distributed. In comparison, the distribution for sandstone is more elongated and slightly left-skewed, with a longer lower tail, indicating mild negative skewness and a slightly higher proportion of mid-to-low-value samples.

In summary, Figure 7 highlights significant differences in feature distribution shapes, central tendencies, and dispersions among the four categories.

To verify the differences in discriminative feature characteristics between tasks within the two-stage framework, this study conducted a feature importance analysis for the shallow-level three-class classification (conglomerate, mud–sandstone, and vegetation) and the deep-level binary classification (mudstone vs. sandstone), as illustrated in Figure 8.

Figure 8 presents the feature importance rankings for different classification tasks within the two-stage random forest framework. The horizontal axis represents the Gini importance scores, where higher values indicate a greater contribution of a feature to classification. The vertical axis lists the 25 feature parameters involved in the analysis. The bar chart on the left (in green) corresponds to the shallow-level three-class classification task, which distinguishes conglomerate, mud–sandstone, and vegetation classes. The features are ranked in descending order of importance. The bar chart on the right (in red) corresponds to the deep-level binary classification task aimed at differentiating between mudstone and sandstone.

As shown in the figure, the same feature contributes differently across classification tasks. In the shallow-level classification, the minimum reflectance exhibits the highest importance, followed by the standard deviation and planarity. In contrast, for the deep-level binary classification, the median and mean reflectance are among the most influential features. This observation underscores the importance of selecting the most discriminative subset of features tailored to the specific classification task to improve model performance.

4.3. Ablation Study of Key Modules

To systematically evaluate the impact of the SG-RFGeo proposed geometry–spectral hybrid features, double-layer random forest classification architecture, and stratigraphic attitude-based post-processing on lithological classification, ablation experiments were conducted using consistent baseline data and evaluation metrics. Key modules were progressively removed to construct four comparative schemes, enabling a quantitative assessment of the contribution of each component. The detailed experimental setup is presented as follows:

(1)

Using only reflectance statistical features with a single-layer random forest classifier;

(2)

Using only geometric features with a single-layer random forest classifier;

(3)

Combining reflectance statistical and geometric features with a single-layer random forest classifier;

(4)

Combining reflectance statistical and geometric features with a double-layer random forest classifier without applying stratigraphic attitude constraints;

(5)

Combining reflectance statistical and geometric features with a double-layer random forest classifier and incorporating stratigraphic attitude constraints (i.e., SG-RFGeo).

The lithological classification results for each experimental scheme, along with comparisons to the manually interpreted ground truth, are presented in Figure 9. A color-coded legend is used to denote the four target classes: conglomerate (orange), sandstone (yellow), mudstone (gray), and vegetation (green), providing an intuitive visual assessment of the classification performance across the different ablation schemes. The manually interpreted ground truth map (artificial result) shown in Figure 9 was constructed based on the field observations and lithological identifications presented in Figure 1, ensuring that the validation reflects the actual geological conditions of the Yueyawan outcrop.

As illustrated in Figure 9, Ablation Study (1) effectively identified extensive vegetation regions, and the banded conglomerate layers located in the lower-middle part of the outcrop were also accurately classified. However, substantial confusion was observed at the mudstone–sandstone boundaries, and the stratigraphic interfaces were not clearly delineated. This suggests that although spectral reflectance statistical features alone can effectively distinguish vegetation from the conglomerate, they are insufficient for accurately differentiating between fine-grained lithologies.

In the results of Ablation Study (2), large areas of vegetation were misclassified as conglomerates, and the overall performance in distinguishing mudstone and sandstone remained unsatisfactory. Nevertheless, a slight improvement was noted in the stratified and continuous regions along the mudstone–sandstone boundary. This outcome is attributed to the direct use of all 17 geometric features without prior selection, which likely introduced redundant or irrelevant features that degraded model performance.

The results of Ablation Study (3) show that among the four target classes, vegetation and conglomerate exhibit high separability, which is largely consistent with the outcomes of the artificial result. However, considerable misclassification remains in the interbedded sandstone–mudstone zones, where the model fails to accurately characterize the distribution of sandstone and mudstone.

Ablation Study (4) showed a marked improvement in classification accuracy across all lithological categories, with a more accurate representation of the spatial distribution. However, limitations persisted in terms of stratigraphic continuity.

In contrast, Scheme V, which incorporates post-processing based on stratigraphic attitude constraints, exhibited the highest consistency with the ground truth provided by expert manual interpretation. This scheme yielded the most accurate representation of lithological distribution in the study area.

To quantitatively assess model performance, the accuracy evaluation framework described in Section 4.1 was applied. The results are summarized in the accuracy evaluation table (

Table 7. Ablation experiment accuracy table.

	OA	Macro Precision	Macro Recall	Macro F1-Score
I	0.632	0.655	0.648	0.650
II	0.655	0.662	0.662	0.658
III	0.805	0.823	0.819	0.821
IV	0.868	0.872	0.869	0.869
V	0.930	0.916	0.931	0.923

).

Table 7 summarizes the overall classification performance of the five experimental schemes. The results reveal a clear trend: using a single feature set—whether spectral or geometric—leads to limited performance, with overall accuracy (OA) below 0.66. Their Macro precision and recall values are very close, indicating a relatively neutral trade-off between false positives and false negatives. This suggests that neither feature type alone is sufficient for robust lithological discrimination in complex or overlapping regions.

Scheme III, which combines spectral reflectance and geometric features within a single-layer random forest model, exhibits a marked improvement over Schemes I and II, with an OA of 0.805 and a Macro F1-score of 0.821. This result demonstrates the effectiveness of feature-level fusion in enhancing class separability. However, some limitations remain, particularly in distinguishing lithologies with subtle spectral and geometric variations (e.g., interbedded sandstone and mudstone).

In contrast, Scheme IV, which employs a double-layer random forest model with the same feature combination, achieves further performance gains across all metrics, with both OA and Macro F1-score exceeding 0.86. Finally, Scheme V, representing the complete SG-RFGeo framework proposed in this study, delivers the best overall performance, achieving an OA of 0.930 and a Macro F1-score of 0.923. These results confirm that the integration of stratigraphic constraints further enhances classification accuracy, promotes geological consistency, and improves balanced recognition across lithological classes.

To further evaluate the classification performance at the individual lithological class level, Figure 10 presents the confusion matrices of all five schemes, expressed in terms of per-class recall. This representation emphasizes the model’s ability to correctly identify each class without being affected by class imbalance. All values in the matrices correspond to recall rates for each lithology.

Schemes I and II, which rely solely on reflectance and geometric features, respectively, exhibit substantial inter-class confusion. In Scheme I, the recall rates for sandstone and mudstone are particularly low, at 52.73% and 61.22%, respectively, with over 30% of sandstone samples misclassified as mudstone. This indicates strong spectral similarity and overlap between these two classes. Scheme II shows a similar issue, with sandstone and mudstone recall rates of 60.00% and 63.27%, and vegetation frequently misclassified as conglomerate (31.25% confusion). Scheme III significantly improved the accuracy of all categories by integrating reflectance features with geometric features, particularly for conglomerate and vegetation, achieving recall rates as high as 86.84% and 90.62%, respectively. The recall rates for mudstone and sandstone also exceeded 70%. However, 21.82% of sandstone samples were still misclassified as mudstone, indicating persistent confusion between the two. Scheme IV adopted the same feature combination and incorporated a double-layer random forest model to further refine the distinction between mudstone and sandstone. This approach achieved an 80% recall rate for sandstone, with over 95% of mudstone correctly classified. Scheme V, representing the SG-RFGeo framework, achieved the most consistent and accurate classification results. Conglomerate, sandstone, and mudstone all exhibit recall rates exceeding 90% (94.74%, 92.73%, and 91.84%, respectively), with negligible inter-class misclassification.

4.4. Comparative Analysis and Performance Evaluation

To further validate the effectiveness of the SG-RFGeo method for lithological classification of outcrop point clouds, comparative experiments were conducted using the Yueyawan outcrop dataset from the Junggar Basin. Four representative classification approaches were selected for comparison: the classical machine learning algorithms Support Vector Machine (SVM) and XGBoost, the deep learning model PointNet, and the unsupervised clustering algorithm K-means. SVM is well-suited for small-sample, high-dimensional data scenarios and has demonstrated strong generalization capabilities in geological classification tasks. XGBoost, an ensemble learning algorithm, enhances classification performance through a gradient-boosting framework. PointNet, a deep learning model designed explicitly for 3D point cloud processing, employs a symmetric function architecture that enables the effective extraction of lithological features under complex terrain conditions. K-means, based on distance metrics, identifies spatial distribution patterns of lithologies by partitioning data into a predefined number of clusters.

The parameter settings for the comparison algorithms are as follows: The SVM model uses an optimized version with a radial basis function (RBF) kernel. The penalty parameter C is set to 1, and the RBF kernel parameter γ is determined via grid search within the range [0.001, 0.1, 1, 10, 100]. Hyperparameter optimization is performed using grid search combined with five-fold cross-validation. The PointNet model retains the standard network structure, with the Adam optimizer used for parameter updates. The learning rate is set to 0.001, the batch size is 32, and the number of training epochs is 100. The loss function is cross-entropy loss. The K-means algorithm uses Euclidean distance as the distance measure, with the ε-neighborhood radius set between 0.5 and 2.0 m. The minimum number of samples (MinPts) is set to 5, and the maximum number of clusters is set to 4 (corresponding to four classes: conglomerate, mudstone, sandstone, and vegetation). The initialization method used is K-means++ to improve clustering stability. The learning rate for XGBoost is set to 0.1, the number of trees is set between 100 and 500, the maximum tree depth ranges from 6 to 10, the subsample ratio is 0.8, the minimum child weight (min_child_weight) is set to 5, and the regularization parameters (λ and α) are set to 0.1 to control model complexity and prevent overfitting.

All comparative experiments were conducted using consistent input data and evaluation metrics to ensure uniform experimental conditions. The classification results of each method are presented in Figure 11.

As shown in Figure 11, the K-means clustering results exhibit significant transitional smoothing and category confusion. The mud–sand interbedding zones are misclassified as large mudstone areas, failing to effectively capture the actual stratigraphic layering structure. PointNet, as a deep learning-based point cloud classification model, shows improvements in the overall recognition of stratigraphic structures; however, its performance in distinguishing between vegetation and gravel is suboptimal. While the SVM algorithm can differentiate between gravel and vegetation, its results lack continuity, and multiple stratigraphic layers are incorrectly split, failing to accurately represent the real continuous layers. Additionally, XGBoost also shows considerable misclassification results for gravel and mudstone. SG-RFGeo demonstrates superior performance in both recognition accuracy and stratigraphic continuity, closely aligning with the expert-labeled bedding, and meets the application requirements for outcrop point cloud lithology classification. The quantitative accuracy evaluation table is shown in

Table 8. Comparison experiment accuracy table.

	OA	Macro Precision	Macro Recall	Macro F1-Score
K-means	0.449	0.443	0.440	0.439
PointNet	0.753	0.752	0.768	0.752
SVM	0.592	0.601	0.614	0.606
XGBoost	0.678	0.732	0.686	0.704
SG-RFGeo	0.930	0.916	0.931	0.923

.

Table 8 presents the overall classification performance of five methods. The SG-RFGeo method achieves the highest accuracy across all metrics, with an overall accuracy (OA) of 0.930 and a Macro F1-score of 0.923, indicating both high correctness and class-level balance. Among the baseline methods, PointNet shows the best performance, with an OA of 0.753 and a Macro F1-score of 0.752, reflecting its strength in handling unstructured 3D point cloud data. XGBoost follows, achieving an OA of 0.678, though its Macro precision (0.732) is higher than its recall (0.686), suggesting it makes fewer false positives but may miss certain classes. SVM performs moderately, with a relatively low OA (0.592) and imbalance between precision and recall. K-means, as an unsupervised method, performs the worst, with an OA of 0.449 and a Macro F1-score of 0.439, indicating poor class separation and limited applicability in this context. Overall, the SG-RFGeo method significantly outperforms both traditional machine learning and deep learning baselines, demonstrating its robustness and suitability for lithological classification in complex outcrop environments.

5. Discussion

5.1. SG-RFGeo Effectiveness Analysis

To address the challenges of automating and improving the quantification of lithological variation in geological outcrop point clouds, this study presents SG-RFGeo, a structured classification framework that integrates spectral and geometric features within a grid-based segmentation scheme. By employing a two-layer random forest classifier and incorporating stratigraphic attitude constraints in the post-processing stage, SG-RFGeo enhances both classification accuracy and geological consistency. The experimental results demonstrate that this integrated strategy effectively reduces misclassification in thinly interbedded zones and improves spatial coherence, particularly in lithologically complex sections.

Spectral features are sensitive to mineral composition but are easily affected by external environmental factors (e.g., weathering of rock surfaces), leading to the phenomenon of “same material, different spectra.” Geometric features, on the other hand, are sensitive to the surface structure of rock masses but cannot distinguish lithologies with similar compositions. Experimental results show that when spectral or geometric features are used individually for lithology classification, the classification accuracy is relatively low, especially when distinguishing between three important lithologies, leading to confusion and failing to reflect the true lithological distribution of the region. The fusion of spectral and geometric features allows for complementary advantages, characterizing outcrop lithology from multiple perspectives, such as geometric morphology and physicochemical properties. As a result, the model’s classification performance is significantly enhanced, with the Macro F1-score increasing from 0.65 to 0.82.

Based on the results of the ablation experiment and expert interpretation, it is demonstrated that the post-processing method incorporating stratigraphic development features further improves the model’s recognition accuracy and aligns with the actual lithological distribution of the study area outcrops. Without the introduction of stratigraphic boundary constraints, the identification of sandstone and mudstone in thin interbedded regions was patchy (Figure 9—Ablation Study (4)), with misclassified areas due to factors such as weathering of the outcrop surface and point cloud scanning quality. While the method partially reflects the lithological distribution pattern of the outcrop, it does not fully meet the requirements for geological applications. By fitting stratigraphic planes and incorporating geological prior information, SG-RFGeo optimizes the lithology boundaries and stratigraphic sequence of the classification results, yielding geological interpretations that closely match the actual elevation and better conform to the lateral continuity of sedimentary strata.

Compared with conventional machine learning algorithms and the deep learning-based PointNet model, SG-RFGeo demonstrates superior classification accuracy. The unsupervised K-means clustering algorithm yielded the lowest accuracy, exhibiting poor performance in distinguishing sandstone–mudstone interbedded regions in the upper portion of the outcrop. Supervised methods such as SVM and XGBoost performed slightly better than K-means, providing a general representation of the lithological distribution but failing to achieve precise boundary delineation. The PointNet algorithm, leveraging a more complex neural network architecture, achieved an accuracy exceeding 75%, representing a notable improvement over SVM and XGBoost, particularly in terms of lithological boundary recognition, as evidenced by the visual results. In comparison, the SG-RFGeo method produced the most accurate classification outcomes, effectively mitigating the adverse effects of surface weathering and vegetation interference. However, the relatively low accuracies of the conventional machine learning models and PointNet are primarily due to the limited discriminative power of point-level XYZ coordinates and raw reflectance values, which are insufficient to distinguish subtle differences between mudstone and sandstone in thin interbedded sequences. Additionally, these algorithms do not incorporate geological prior knowledge, which is essential for accurate lithological classification in complex outcrop settings. Furthermore, the limited number of labeled samples in thin beds likely constrained the performance of PointNet, which typically requires large datasets to fully leverage its deep learning architecture.

5.2. Analysis of Key Spectral–Geometric Features

The overall feature importance ranking chart (Figure 6) shows that the standard deviation, minimum value, and median in the spectral statistical features exhibit the highest discriminative power, with the standard deviation and minimum value being significantly more important than the other features. The standard deviation reflects the local heterogeneity of mineral composition, while the minimum value is related to the primary mineral components of the rock and the absorption wavelengths. As shown in the violin plot (Figure 7), a clear distinction exists between the plant and sedimentary rock categories, with features such as Std Dev (standard deviation) and Min Value (minimum value) showing a broad and dispersed distribution, indicating a wider coverage in the plant category. High standard deviation and low flatness are attributed to the complex surface structures of vegetation, including leaves, branches, and voids. The minimum value corresponds to the strong absorption characteristics in the near-infrared wavelengths of vegetation (e.g., 1400–2500 nm), which are closely related to the absorption peaks of chlorophyll and water content.

Conglomerates, due to the coarse-grained mixed accumulation (>2 mm quartz minerals, flint clasts) and weak cementation, have rough surfaces and heterogeneous composition, resulting in a higher standard deviation of reflectance compared to sandstone and mudstone, as well as lower flatness. In contrast, the environmental deposition of sandstone and mudstone (lacustrine or shallow marine) is lower energy than conglomerates, with stronger diagenetic compaction, leading to a higher level of particle homogenization. This results in more stable reflectance and higher flatness.

From the violin plot of the median values, it can be observed that the distribution of mudstone follows a symmetric bell-shaped curve, which is more concentrated than sandstone, reflecting the differences in rock composition between the two. Sandstone is primarily composed of medium- to fine-grained particles with strong cementation, resulting in a relatively stable surface structure and more uniform reflectance variation. However, due to changes in the depositional environment or secondary cementation, the porosity and cementation degree of sandstone vary, leading to fluctuations in the reflectance median within a certain range. Mudstone, on the other hand, is mainly composed of very fine clay minerals with smaller particle sizes and oriented arrangement, exhibiting higher compactness and more stable reflectance with a more concentrated median distribution.

The above analysis indicates that the spectral and geometric features developed in this study effectively reveal the physicochemical properties of different sedimentary rocks, aligning with their depositional environments and diagenetic evolution patterns.

5.3. Grid Resolution Effect Analysis

To investigate how grid resolution affects classification performance, comparative experiments were conducted with grid cell sizes of 5 mm, 10 mm, and 20 mm. Local comparison results of these experiments are illustrated in Figure 12.

As shown in Figure 12, when the grid size is reduced to 5 mm (Figure 12a), the classification erroneously displays interbedded patterns in areas that are not actually interbedded, with sections of conglomerate misclassified as alternating sandstone and conglomerate layers, and the underlying sandstone regions incorrectly identified as mixed mudstone–sandstone interbeds. In contrast, the coarser 20 mm grid (Figure 12c) leads to excessive smoothing, failing to detect the sandstone layers altogether and overestimating the extent of conglomerate, with thin interbeds being merged or lost, resulting in a loss of boundary precision. The 10 mm grid (Figure 12b) provides a balanced result, preserving clear lithological boundaries while avoiding excessive noise, and demonstrates good agreement with the manually interpreted ground truth (Figure 12d). The detailed accuracy metrics are presented in

Table 9. Classification accuracy for different grid resolutions.

	OA	Macro Precision	Macro Recall	Macro F1-Score
5 mm	0.578	0.592	0.601	0.590
10 mm	0.930	0.916	0.931	0.923
20 mm	0.507	0.351	0.551	0.426

.

As shown in Table 9, the 10 mm grid resolution achieved the highest OA of 0.930, demonstrating its superiority in balancing spatial detail and feature stability. In contrast, the 5 mm grid, despite offering finer spatial resolution, suffered from insufficient point counts per cell, which limited feature representativeness and reduced feature stability, ultimately leading to increased misclassification and lower overall accuracy (0.578). The 20 mm grid, while aggregating more points per cell, overly smoothed lithological boundaries, resulting in the lowest accuracy (0.507) and the poorest performance across all metrics. These results confirm that a 10 mm grid effectively captures lithological variability while maintaining robust feature extraction, underscoring its suitability for lithological classification in thinly interbedded and weathered outcrop settings.

5.4. Case Study on Carbonate Outcrop

To further assess the applicability of SG-RFGeo, we conducted a case study on a carbonate outcrop in Sichuan, which includes silty crystalline dolomite, algal dolomite, and argillaceous shale (Figure 13a). Point cloud data for the carbonate outcrop were acquired using the same terrestrial laser scanning (TLS) system, with a minimum point spacing of 1 mm, and processed following the identical regular grid-based segmentation and feature extraction procedures described earlier. The same spectral–statistical and geometric feature sets were employed, and the double-layer random forest classifier was retrained using manually labeled samples representative of the carbonate and shale lithologies. Specifically, the first stage separated argillaceous shale from the carbonate lithologies, while the second stage distinguished algal dolomite from silty crystalline dolomite within the carbonate group. The experimental results are shown in Figure 13c.

By comparing Figure 13b and Figure 13c, it can be observed that the lithological distribution identified by SG-RFGeo is largely consistent with the field validation results. The left part of the outcrop exhibits interbedded layers of silty crystalline dolomite and algal dolomite, the central section is dominated by silty crystalline dolomite, and gray argillaceous shale appears on the right side. Detailed quantitative evaluation metrics are presented in

Table 10. Classification performance on the carbonate outcrop case study.

	Precision	Recall	F1-Score	OA
algal dolomite	0.735	0.818	0.774	0.861
silty crystalline dolomite	0.948	0.853	0.898
argillaceous shale	0.767	0.943	0.846

.

As summarized in Table 10, the SG-RFGeo framework achieved an OA of 86.1% for the carbonate outcrop case study. Among the three lithologies, silty crystalline dolomite exhibited the highest precision (0.948) and strong overall performance, with an F1-score of 0.898; however, this may partially reflect the dominance of this lithology in the dataset, which can positively bias performance metrics for majority classes. Argillaceous shale achieved a high recall of 0.943, indicating the model’s effective detection of this minority lithology, due to its homogeneous composition, which leads to more consistent spectral features for classification. Algal dolomite showed a moderate performance, with an F1-score of 0.774, primarily limited by lower precision (0.735) compared to the other classes. These results demonstrate that SG-RFGeo maintains robust and balanced classification capability across different carbonate lithologies, effectively distinguishing between silty crystalline dolomite, algal dolomite, and argillaceous shale in complex outcrop environments.

6. Conclusions

This study presents SG-RFGeo, a rapid and effective method for the lithological classification of complex geological outcrops. The massive outcrop point cloud data are segmented into smaller subsets using a regular grid-based partitioning strategy, from which spectral–statistical and geometric features are extracted. A two-layer random forest classifier is then employed for lithology identification, with geological prior knowledge—specifically stratigraphic orientation—incorporated into the post-processing stage to enhance classification accuracy.

Based on the experimental outcomes, the following conclusions can be drawn:

The spectral–geometric features derived from regular grid segmentation effectively capture lithological variations among different rock types, and when combined with the two-stage random forest architecture, enable refined discrimination of easily confused lithologies.

Incorporating geological knowledge, such as stratigraphic orientation, into the classification process significantly enhances the geological plausibility of the results and improves the overall model performance.

Compared with classical classification methods, the proposed approach achieves substantially higher accuracy, with an overall classification accuracy of 93%, satisfying the practical requirements of geological exploration.

These findings confirm that by integrating multimodal spectral–geometric features, applying a hierarchical classification strategy, and incorporating geological prior constraints, SG-RFGeo effectively addresses the challenges of thin interbedded lithologies and weathered surfaces, leading to significantly improved classification precision and geologically coherent results in complex outcrop settings.

Additionally, we conducted preliminary experiments applying the SG-RFGeo framework to carbonate outcrops containing lithologies such as silty crystalline dolomite and algal dolomite, along with argillaceous shale, which demonstrated promising classification performance. These results suggest that the proposed method has potential applicability to a wider range of rock types and geological environments beyond the clastic sediments investigated in this study. Despite these encouraging outcomes, several limitations remain, particularly in data acquisition. The performance of terrestrial laser scanning (TLS) systems is highly dependent on environmental conditions, such as weather and visibility, which can affect the quality of the acquired data. Future work will explore integrating complementary remote sensing techniques, such as radar or geophysical surveys, to mitigate the limitations of TLS under adverse weather conditions. Furthermore, upcoming research will focus on interdisciplinary integration and the continuous refinement of stratigraphic constraint strategies to further enhance classification accuracy in complex geological settings, particularly in scenarios involving stratigraphic pinch-outs, intrusive contacts, and tectonic overprinting, which are indeed important directions for future research.