The study area is located around Daxi Town, Pinghe County, Zhangzhou City, Fujian Province. The specific administrative division map, true-color composite image, and DEM are shown in Figure 1 . It is situated in a complex region with intersecting Nanshan Mountain Range and the Second Complex Uplift Belt of the Xinhua-Xia series. It has undergone multiple crustal movements and has exhibited complex tectonic faults. The primary geological formations consist of Lower Cretaceous, Jurassic strata, and Yanshanian intrusion rocks. The main rock types in this area include Rhyolite tuff (RhyT), Yingan tuff (YinT), Rhyolite (Rhy), sandstone (San), and granite (Gra). Sedimentary deposits consist mainly of conglomerate, gravel, and sandy soil layers (CGS).

The area is characterized by dense vegetation growth, with a forest coverage rate of 73.2% as of 2021. It exhibits a variety of plant community types and a complex hierarchical structure, resembling the vibrant landscapes of the South Asian tropical rainforests. The main vegetation types include coniferous forests, broad-leaved forests, mixed forests of conifers and broad-leaved trees, bamboo forests, shrubs, and grass slopes. The bedrock is mostly covered by a few outcrops (Shi and Wang, 2014). Additionally, in highly vegetated areas, the surface layer is heavily weathered, and the soil layer is thick, posing challenges for rock-type identification using remote sensing techniques.

The Sentinel satellite carries the Multispectral Instrument (MSI), capturing imagery data in visible, near-infrared, and shortwave infrared bands. It provides a spatial resolution ranging from 10 m to 60 m (band details in Table 1 ). With a 5-day revisit period at the equator, it covers a swath width of 290 km (Chen et al., 2021). The S2 MSI Level-2A products offer preprocessed bottom-of-atmosphere reflectance images, including geometric correction, orthorectification, image registration, radiometric calibration, and atmospheric correction. A total of 78 images for the year 2021 with cloud coverage of less than 10% in the study area were obtained, and cloud masking using the QA60 band was performed to eliminate cloud effects. Bands B5, B6, B7, B8A, B11, and B12 were resampled to 10-meter spatial resolution using bilinear interpolation, and then the median composite was applied to the images.

The Shuttle Radar Topography Mission Digital Elevation Model (SRTM DEM) is a digital representation of terrain elevation obtained using radar equipment carried by space shuttles, covering over 80% of the Earth’s land surface. SRTM DEM data is acquired by emitting radar beams from a space shuttle towards the Earth’s surface and measuring the returning signals. This technique bypasses cloud cover and vegetation obstruction to acquire surface elevation data. The SRTM DEM 30 m dataset was resampled to 10 m using bilinear interpolation to match the spatial resolution of the optical imagery.

The reference data includes the regional geological map from the First Survey Team of Fujian Provincial Geological Survey Bureau in 1982 and lithology distribution maps obtained through a combination of vegetation suppression methods and manual visual interpretation by professional experts. We adopted the approach used in the lithological distribution map to classify land surface cover types and collected pixel-level sample data for 7 land cover classes based on the regional geological map: CGS (118), San (106), Rhy (112), RhyT (121), YinT (121), Gra (110), and water (39).

The main idea for pixel-wise rock classification is based on “geobotany”. It uses differences in surface reflectance between vegetation communities to indirectly identify underlying rock types by considering terrain and landform characteristics. To achieve this, we construct five classical machine learning algorithms: minimum distance algorithm (MD), classification and regression trees (CART), support vector machine (SVM), random forest (RF), and gradient boosting decision tree (GBDT) (refer to Figure 2 ).

The following research plan aims to reduce computational burden and accelerate the production of optimal lithological spatial distribution maps. The available geological data is divided into training and validation sets at a 7:3 ratio to assess the model’s generalization ability, which evaluates its performance on new data. Then, five classification algorithms are trained and optimized using the preprocessed Sentinel-2 remote sensing image and a sample dataset. The optimized models are validated and evaluated for accuracy using validation samples. Subsequently, the optimal model is used to classify rock types and generate a lithological map based on Sentinel-2 data. In the subsequent studies based on SRTM DEM (alone) and combination of Sentinel-2 with SRTM DEM, at least two classifiers that performed well in the previous step are used for validation and accuracy evaluation. Similarly, the optimal model is used to generate lithological maps based on SRTM DEM (alone) and combined with Sentinel-2 and SRTM data.

B2, B3, B4, B5, B6, B7, B8, B8A, B11, and B12, were selected from Sentinel-2. It should be noted that B1 represents the aerosol band and B9 represents the water vapor band, therefore, they were not used in this study. Additionally, five spectral features were computed, which are crucial for rock or water body identification: enhanced vegetation index (EVI), green chlorophyll vegetation index (GCVI), normalized difference building index (NDBI), normalized difference water index (NDWI), and Land Surface Water Index (LSWI). This study employed EVI to indicate vegetation growth status, as it is well known that normalized difference vegetation index (NDVI) can saturate areas with high vegetation cover (Huete et al., 1997). GCVI is sensitive to chlorophyll and can be used to identify agricultural areas (Huete et al., 2002). NDBI is useful for identifying built-up areas (Benbahria et al., 2018). NDWI and LSWI are vegetation indices that are highly sensitive to surface water (Jeong et al., 2012).

For SRTM DEM, six terrain features, including elevation (E), slope, aspect, vertical curvature, horizontal curvature, and Gaussian curvature, were obtained using terrain analysis algorithms proposed by Florinsky (2016) and Safanelli (Safanelli et al., 2020). Additionally, surface roughness (SR), high integral (HI), and surface index (SI) were calculated within a 3 × 3 window, providing indications of terrain features associated with rock type information.

Where E is extracted directly from the DEM. E_mean, E_max, and E_min represent the mean, maximum, and minimum values within the moving window, respectively. SR is the ratio of grid surface area to projected area, which is used to quantify tectonic landform changes. A higher SR value indicates a more severe degree of regional deformation (Han et al., 2021). SI is a new efficient index that simultaneously depicts the preserved and eroded portions of the landscape (Andreani et al., 2014).

The Minimum Distance (MD) algorithm is widely used for classification and pattern recognition (Wacker and Landgrebe, 1972). It assigns input data to the closest category based on computed distances between categories, using methods such as Euclidean, Manhattan, and Minkowski distances. In the context of lithological mapping, this algorithm can be used to infer the lithology of specific locations based on their proximity to known geological features or data points (Pal et al., 2020). MD is simple, easy to implement, and well-suited for problems with distinct category boundaries.

Classification and regression trees (CART) is a tree-based algorithm for classification and regression (Friedl and Brodley, 1997; Pal and Mather, 2003). It recursively splits and evaluates input data to build a tree-like model, where nodes represent features, branches represent feature values, and leaf nodes represent final results. Applying CART to lithological mapping involves systematic interpretation of geological features for accurate lithological classification. Visualizing the decision tree offers insights into hierarchical feature divisions that lead to lithological categorization, enhancing our understanding of how different features influence prediction accuracy (Serbouti et al., 2022). CART has a simple structure, making it easy to understand, interpret, and generate decision-making rules. It’s worth noting that the effectiveness of the CART decision tree method relies on the quality and relevance of selected features, as well as the representation of distinct lithological classes in the dataset (Lewis, 2000).

Support vector machine (SVM) is a non-parametric classifier widely used for binary and multi-class classification tasks (Pal and Mather, 2005; Mountrakis et al., 2011). It aims to minimize structural risk by identifying an optimal hyperplane in the feature space that maximizes the margin between samples of different classes. SVM can be used for lithologic classification by selecting training and testing pixels, training the SVM classifier using a kernel function, and mapping the input data (such as spectra, textures, topography, gamma-ray and land temperature) (Othman and Gloaguen, 2014; Harris and Grunsky, 2015; Yang et al., 2018; Chen C. et al., 2023) into a higher-dimensional space to find a hyperplane that separates the different classes. SVM has the advantage of handling high-dimensional data, nonlinear classification problems, and small sample sizes (Shebl and Csamer, 2021; Shebl et al., 2021). However, it also has limitations in terms of selecting appropriate kernel functions, computational intensity, and sensitivity to noise and outliers (Othman and Gloaguen, 2017).

Random forest (RF) works by creating multiple decision trees on randomly selected subsets of the data and then combining the results to make a final prediction (Pal, 2005; Belgiu and Drăguţ, 2016). It has significant advantages with the ability to combine multiple remote sensing and data sources in lithology mapping as it improves its generalization ability by randomly selecting input or input combinations at each node (Breiman, 2001). It is especially effective for processing high-dimensional and noisy input data and can overcome the interference of vegetation coverage, thereby improving the accuracy of lithological mapping (Harris and Grunsky, 2015; Bachri et al., 2019). However, caution should be exercised when fine-tuning parameters for optimal outcomes and effectively managing computational expenses, particularly when dealing with substantial datasets. The risk of overfitting due to an abundance of trees or noisy data should be considered, along with its limited efficacy with imbalanced datasets (Guo et al., 2022).

The Gradient Boosting Decision Tree (GBDT) is an iterative ensemble learning algorithm that constructs a strong prediction model (Yang et al., 2018; Xu et al., 2020). It builds multiple decision trees iteratively, using the residual between the current predicted value and the true label to train each tree. Each iteration adjusts the predicted value to approximate the true label. The final prediction is obtained by combining the predictions of all the trees. Although the algorithm is rarely used in lithology mapping in high vegetation cover areas, studies have shown that it is effective in handling large datasets and high-dimensional feature spaces, and it is robust to noise, outliers, and missing data (Lemercier et al., 2012; Zhou et al., 2020; Cai et al., 2022).

Accuracy assessment objectively evaluates the performance of remote sensing algorithms and models, providing a reliable foundation for remote sensing applications and decision-making (Hay, 1988). It validates the accuracy of tasks like image classification and object detection, helping to determine the reliability and feasibility of the results. The Confusion Matrix is a tabular representation used to evaluate the performance of a classification model (Comber et al., 2012; Salmon et al., 2015). Therefore, we calculate four evaluation metrics using the Confusion Matrix to assess the classification results and optimize the model: overall accuracy (OA), recall (R), precision (P), and Kappa coefficient.

Model tuning optimizes the performance and generalizability of the machine learning model by adjusting its parameters or hyperparameters, improving its effectiveness in addressing real-world problems. By systematically adjusting parameters, the model can better adapt to data patterns and characteristics. Proper parameter settings also reduce computational resources required for training and prediction. Through tuning, structure of the model and parameter selection are optimized, resulting in reduced computational costs and time consumption.

The sample dataset was randomly divided into training and validation sets at a 7:3 ratio for model training and tuning based on Sentinel-2 imagery. To determine the best settings for our models, we test them by using different parameters in the validation set. We begin by defining a range of parameter values and then try different values within that range. Finally, we select the parameter values that make our models perform the best on the validation set. We can show this process on a chart, where the horizontal line represents different parameter choices, and the vertical line shows the performance of the model (typically, model performance metrics such as OA). The tuning results for each model are shown in Figure 3 . The MD algorithm achieved optimal performance using the Mahalanobis distance metric. The CART was tuned with a maximum of 50 nodes and a minimum of 5 nodes. The gamma coefficient of SVM model was set to 20, and the Cost parameter was set to 10. The RF utilized 210 trees and a minimum of 3 leaf nodes. The GBDT employed 80 trees and a maximum of 70 leaf nodes.

In machine learning and data modeling, the optimal parameters of a model typically change with different datasets and feature combinations. Once we have identified at least two promising classifiers based on Sentinel-2, we apply the same tuning methodology to both the DEM data and the combination of Sentinel-2 and DEM data to ensure optimal model performance.

The model accuracy and validation accuracy of five typical machine learning algorithms were summarized (as shown in Figure 4 ). The SVM algorithm had higher model accuracy but lower validation accuracy, possibly indicating overfitting due to linear inseparability of rock categories and limited sample size. MD and CART performed poorly, with model accuracy below 70% and validation accuracies below 0.4, suggesting limitations in handling complex structural information. RF and GBDT showed better performance as ensemble algorithms with strong learning capabilities with OA 46.82% and 45%, and Kappa 0.371 and 0.349 respectively.

Figure 5 illustrates rock-type accuracy achieved by the five machine learning algorithms. The “water” class consistently exhibits high accuracy (close to 100%) across all models, indicating the effectiveness of the algorithms. In terms of class accuracy, the RF model demonstrates notable improvements in precision for “RhyT”, “San”, and “YinT” classes, as well as increased recall for “RhyT” and “YinT” classes. For instance, precision of San improved by 13.1% to 20%, and recall of “RhyT” increased by 10% to 36.6%. The GBDT achieved overall high accuracy through significant improvements in precision and recall for “CGS” and “San”. Specifically, precision and recall of “CGS” increased by 1.3% to 11.8% and 2.7% to 29%, respectively.

Using RF and GBDT algorithms for lithology classification achieves higher accuracy based on the accuracy assessment results (see Figures 6 , 7 ). The overall classification accuracy is approximately 49% with a kappa coefficient of approximately 0.4. When comparing individual rock types, both algorithms show higher accuracy for “CGS” and “Rhy”. “CGS” plays a significant role in terrain morphology and exhibits a certain relationship with terrain features. “Rhy” is commonly found in uplift zones associated with tectonic landforms.

The lithology classification map generated using SRTM DEM (Digital Elevation Model) data shows higher accuracy than Sentinel-2 imagery. The overall improvement in lithology classification accuracy is mainly attributed to better classification of “Rhy” and “San”, despite a slight decrease in the accuracy of classifying “RhyT” and “Gra”. This can be explained by the strong correlation between spatial distribution and landform morphology. “San” is commonly found in low-lying areas, while “Rhy” is associated with volcanic activity and its distribution relates to volcanic topography. However, it should be noted that the accuracy of “water” based on SRTM DEM is significantly lower than that of Sentinel-2 imagery due to the limited information about water bodies in DEM, which primarily focuses on surface terrain elevation. As a result, there may be errors or omissions in the parts of the lithology classification that involve water bodies, as it does not capture the characteristics of water flow from higher to lower elevations. Overall, the utilization of DEM improves the accuracy of lithological classification by capturing the relationship between lithological variations and landform morphology.

The RF and GBDT are also utilized in a lithology mapping study that combines Sentinel-2 imagery and SRTM DEM data. The extraction of lithological information can be greatly improved by integrating optical remote sensing data with terrain morphology features, resulting in more comprehensive and accurate spatial information. Both algorithms (RF and GBDT) show a significant improvement in accuracy compared to Sentinel-2 (alone) or SRTM DEM (alone), achieving OA of approximately 60% (see Figure 6 ). The GBDT demonstrates a more precise lithology classification, achieving OA of 63.18% and a Kappa of 0.565. Compared with RF, GBDT improved OA and Kappa by 4.54% and 0.053, respectively.

The joint utilization of Sentinel-2 imagery and SRTM DEM, along with advanced machine learning algorithms like GBDT, significantly improves the classification accuracy of various lithological categories, as depicted in Figure 7 . The performance in classifying “CGS”, “San” and “Rhy” is notably enhanced. Moreover, the classification of “water” achieves optimal results.

Based on the above study, lithological mapping was performed using RF for Sentinel-2 (alone), GBDT for DEM (alone) and the combination of Sentinel-2 and DEM ( Figure 8 ). In the lithological map based on Sentinel-2, the limited distinguishability of rock types other than “CGS” and the presence of significant “speckle” phenomenon may be attributed to the optical sensor limitations of Sentinel-2, which are affected by cloud cover, atmospheric interference, and vegetation obstruction. In contrast, DEM data offers continuous surface elevation and terrain information, resulting in a clearer representation of lithological distribution. The combination of Sentinel-2 imagery and DEM allows for leveraging their respective strengths, leading to more comprehensive and accurate rock-type classification results.