Hyperspectral Rock Classification Method Based on Spatial-Spectral Multidimensional Feature Fusion

Abstract

The issues of the same material with different spectra and the same spectra for different materials pose challenges in hyperspectral rock classification. This paper proposes a multidimensional feature network based on 2-D convolutional neural networks (2-D CNNs) and recurrent neural networks (RNNs) for achieving deep combined extraction and fusion of spatial information, such as the rock shape and texture, with spectral information. Experiments are conducted on a hyperspectral rock image dataset obtained by scanning 81 common igneous and metamorphic rock samples using the HySpex hyperspectral sensor imaging system to validate the effectiveness of the proposed network model. The results show that the model achieved an overall classification accuracy of 97.925% and an average classification accuracy of 97.956% on this dataset, surpassing the performances of existing models in the field of rock classification.

1. Introduction

In recent years, hyperspectral rock image classification has been demonstrated to have significant advantages in geological exploration, such as mineral resource localization, geological structures analysis, and environmental effect assessment [1,2,3,4,5]. High-resolution hyperspectral data, derived from satellites, airborne systems, and laboratory settings, have distinct advantages in various application scenarios using advanced artificial intelligence algorithms, such as machine learning and deep learning [6]. Satellite remote sensing images, characterized by a wide coverage and high temporal resolution, have been commonly used to conduct large-scale rock and mineral classifications. For example, multispectral and hyperspectral data, provided by satellites such as Landsat 8 and Sentinel-2, have been widely used in geological surveys [7,8]. Researchers have utilized these data and machine learning algorithms such as support vector machines (SVMs) [9] and random forest (RF) models [10] to precisely classify types of surface rocks. Similarly, airborne hyperspectral remote sensing systems, such as the hydrological modeling and analysis platform (HyMap) and airborne visible/infrared imaging spectrometer (AVIRIS), can acquire high-spectral-resolution data that accurately reflect the spectral characteristics of objects [11,12]. For example, several studies have achieved high-precision mineral composition classification and spatial distribution analysis by combining convolutional neural networks (CNNs) [13] and long short-term memory (LSTM) networks [14] with geographic information system (GIS) technology.

However, hyperspectral rock imagery acquired using satellites and airborne platforms predominantly emphasizes the breadth and richness of geological remote sensing and falls short of meeting the demand for the precise classification of individual rock samples [15,16]. In contrast to conventional satellite and airborne remote sensing, fine spatial resolution imaging spectroscopy coupled with quantitative analysis of geological materials in the laboratory or via controlled indoor environments is gradually becoming a mainstream research method [17,18]. Laboratory imaging spectrometers are capable of conducting advanced imaging spectroscopy on samples, cuttings, and cores, delivering spectral information that mirrors the surface reflectance spectra of the specimens that are characterized by exceptionally high spatial resolution and lateral continuity. This technology not only facilitates the identification of minerals and their mixtures but also evaluates the spatial distribution and variability of these minerals across the sample surface [19]. Consequently, hyperspectral rock data derived from laboratory platforms can be used to validate and calibrate airborne and satellite hyperspectral imagery, offering a more comprehensive evaluation framework for geological research [20].

Currently, as hyperspectral imaging systems become more portable, research on rock classification using hyperspectral rock sample images obtained in the laboratory or via controlled indoor environments is gaining more attention, and several case studies have been published. For example, Guo et al. [21] found that using high-frequency features after discrete wavelet transform (DWT) [22] can improve the rock discrimination accuracy. They conducted classification experiments on hyperspectral images of six types of rock samples, including dolomite, andesite, and tuff, obtained from a laboratory platform combined with a random forest classifier. Their experiments verified that the DWT can capture signal mutations and reveal subtle differences between different rock spectra, improving the overall classification accuracy by 10%–20% compared to that without discrete wavelet transformation. Galdames et al. [23] constructed a hyperspectral rock dataset for 13 types of rocks, including andesite, granodiorite, and breccia, using a laboratory-based hyperspectral sensor. They applied the mask region-based convolutional neural network (R-CNN) framework [24] to dimensionally reduced data through a fully convolutional network for rock classification. This method performed instance segmentation on each rock, achieving an accuracy of over 95% in classifying the 13 types of rock. Hamedianfar et al. [25] proposed using core scanning in the long-wave infrared wavelength region to map hyperspectral images of five types of rock, namely, quartz, talc, feldspar, chlorite, and quartz-carbonate, with a laboratory platform. Using a deep learning framework with the U-net architecture called ENVINet5 [26], they achieved pixel-wise precise classification with an overall accuracy of 82.74%.

Although current research on hyperspectral rock image classification has achieved some results, there are still many shortcomings. In the classification of hyperspectral rock images obtained using satellites and airborne platforms, traditional machine learning methods only extract shallow spectral features and fail to fully exploit the deep feature values of hyperspectral data, making it difficult to accurately identify mixed and overlapping rock types [27,28,29]. In deep learning methods, such as the multilayer perceptron neural network used by Bahrambeygi et al. [30], the network structure ignores the spatial structure of the input data and only performs linear combinations, making it difficult to extract deep spatial information, resulting in a low accuracy for spectrally classifying mixed rocks. In studies conducted in the laboratory or using controlled indoor environments on hyperspectral rock image classification, the spectral features of different rock and mineral samples exhibit many similarities, leading to the issues of the same spectra for different materials and different spectra for the same material, thus resulting in both misclassification and low classification accuracy. For example, although Guo et al. [21] improved the classification accuracy to some extent using DWT, the large number of high-dimensional features generated via DWT were not effectively selected, resulting in a relatively low overall classification accuracy. To address these issues, researchers have introduced a series of new methods. Zhang et al. [31] combined a 1-D CNN with a 2-D CNN, which enhanced the model’s ability to extract rock spectral features and improve the classification accuracy. Galdames et al. [23] used the mask R-CNN framework to conduct instance segmentation of rocks and achieved high accuracy in classifying certain types of rock. Hamedianfar et al. [25] used a deep learning framework with a U-net architecture called ENVINet5 to conduct precise pixel-wise classification, which improved the mineral category classification and captured subtle features. However, these methods still have limitations. For instance, Galdames et al.’s [23] method lacks an explanation of the phenomenon of different spectra for the same material, and Hamedianfar et al.’s [25] method still leaves some pixels unclassified when dealing with certain spectrally mixed categories.

Therefore, in this study, we utilized a push-broom orbit on a laboratory platform combined with the HySpex hyperspectral sensor imaging system to perform hyperspectral imaging on 81 common igneous and metamorphic rock samples, acquiring hyperspectral rock images and establishing ground truth labels. Because the compositional and spectral curve trends of the igneous and metamorphic rock samples had classification characteristics, the 81 rock samples were initially divided into 28 categories. Through spectral analysis of the rock samples, it was found that the spectral curves of the single-point reflections from the rock surface were closely related to the mineral composition of the rock. However, when dealing with complex rock samples with mixed or similar mineral compositions, misclassification often occurs due to overlapping or similar spectral curves. In addressing this issue, in this study, we demonstrated that the synergistic fusion of deep spatial information extracted using a CNN and spectral features can significantly improve the classification accuracy of hyperspectral rock images.

2. Materials and Method

2.1. Rock Dataset

In this study, spectral data of rocks were collected using the HySpex shortwave infrared (SWIR)-384 hyperspectral camera imaging system, manufactured by Norsk Elektro Optikk, located in Oslo, Norway (as shown in Figure 1). And the FieldSpec3 ASD spectroradiometer in a controlled darkroom laboratory environment. The HySpex SWIR-384 hyperspectral camera system covers a spectral range of 950 to 2500 nm, with 288 spectral bands and a resolution of 5.45 nm. The FieldSpec3 ASD spectroradiometer measures the near-infrared region, with a range of 1001 to 2500 nm and a resolution of 8 nm.

The research team conducted scanning and imaging of 81 freshly exposed rock surfaces, establishing a publicly available hyperspectral rock image dataset [32], which includes 35 metamorphic rocks and 46 igneous rocks (as shown in Figure 2a). The effectiveness of the proposed model was validated on this dataset.

In the preliminary phase of the project, 50 random 5 × 5 pixel ROIs were sampled for each rock, and the average spectra were calculated to generate the spectral curves shown in Figure 2b. Based on the overall spectral similarity and local absorption–reflection characteristics of the rocks, a preliminary classification system of 28 rock categories was established [33]. The names of the rocks included in each category are shown in Figure 2c and

Table 1. Names of rocks included in the 28 types of rocks.

Class Name	Rock Name	Class Name	Rock Name
r1	komatiite, chlorite schist	r15	andesite, trachyte, syenite, pitchstone
r2	garnet granulite, amphibolite, gabbro pegmatite, eclogite, amphiboleschis, plagioclase amphibole schist	r16	quartz diorite, orthoclase porphyry, granitic gneiss
r3	finely crystalline marble, mesocrystalline marble, red marble	r17	biotite gneiss, ptygmatite
r4	intrusive carbonate	r18	aplite
r5	garnet skarn	r19	grayish white slate, rutile schist, muscovite quartz schist, staurolite schist
r6	epidote skarn, garnet epidote skarn, phyllite	r20	diorite, granodiorite, nepheline-syenite, nosean phonolite
r7	striped migmatite, augen migmatite	r21	peridotite, diabase, layered magnet quartzite
r8	pegmatite, biotite hornfels, greisen, mylonite, sillimanite schist	r22	amphibole eclogite, leptite
r9	kyanite schist	r23	diorite porphyrite, ijolite, lamprophyre
r10	granite, fine-grained granite, monzonitic granite, porphyritic granite, felsite, alaskite	r24	pyroxene quartz orthoclase porphyry, perlite
r11	k-feldspar granite, quartzite	r25	pyroxenite, gabbro, amygdaloidal basalt
r12	anorthosite, graphic granite, rhyolite, pseudoleucite phonolite	r26	kimberlite, basalt, vesicular basalt, andalusite hornfels, cordierite hornfels, serpentine
r13	lithophysa rhyolite, migmatitic granite	r27	lapilli (slag), black slate
r14	trachyandesite	r28	volcanic lava, obsidian, pumice

.

On this basis, we established ground truth labels for the 28 rock classes (Figure 3a). This facilitated the division of these classes into training, validation, and testing sets for the subsequent supervised classification network models. When generating the ground truth labels for the 28 rock classes, we encountered challenges due to the presence of blurred edges and shadows in the rock samples. To address this issue, a small portion of the data was excluded by trimming the edges of the rock fragments. To achieve this, we utilized image morphological processing techniques and applied a 5 × 5 erosion algorithm to remove the boundary parts of the rock samples (Figure 3b). The colors corresponding to the 28 rock labels and the number of pixel samples are shown in

Table 2. The 28 types of rock labels corresponding to colors and number of pixel samples.

Class	Samples	Class	Samples	Class	Samples	Class	Samples
r1	11,415	r8	22,042	r15	19,547	r22	10,746
r2	30,779	r9	4718	r16	14,048	r23	14,108
r3	14,473	r10	30,620	r17	10,351	r24	9242
r4	5121	r11	9534	r18	5033	r25	14,145
r5	4434	r12	19,353	r19	19,531	r26	26,830
r6	11,490	r13	9590	r20	21,145	r27	10,381
r7	11,129	r14	5471	r21	13,703	r28	11,259
TOTAL				390,238

.

2.2. Methods

2.2.1. Related Work

Despite years of development, significant progress has been made in hyperspectral image classification; however, handling high-dimensional spectral data remains challenging. Deep learning, especially CNNs, has become a focal point in hyperspectral rock image classification. Originally designed for 2-D images, CNNs excel in capturing spatial features [34,35,36] (Figure 4).

The formula for convolution is as follows:

$$F^m=f(F^{m-1}\ast W^m+b^m)$$

(1)

where f(•) is the nonlinear activation function, which enhances the algorithm’s ability to process nonlinear data. $F^{m-1}$ is the input feature map of the (m − 1)th layer, and $F^m$ is the output feature map of the mth layer. $W^m$ is the convolution filter, and $b^m$ is the bias value for each output feature map.

The main objective of the convolution operation is to extract various features from the input data. While some convolution layers excel at capturing low-level features, such as edges, lines, and contours, the deeper convolution layers progressively extract more complex features, such as textures, based on these fundamental features. In the context of rock samples, the convolution layers are particularly effective at discerning spatial information, aiding in the accurate identification of key attributes such as rock locations, edges, and textures. After passing through the convolution layers, the features become high-dimensional. In this stage, pooling layers play a crucial role. These layers segment the features into different regions and compute the maximum or average value within these regions, thereby generating new, lower-dimensional features. Pooling layers are typically placed between consecutive convolution layers to reduce the volume of data and the parameters and prevent overfitting.

In terms of spectral information extraction, we employed recurrent neural networks (RNNs) to extract the spectral information, combined with spatial data extracted using 2-D CNNs, to establish a spectral-spatial fusion network. The specific RNN variant chosen for this purpose was an enhanced model known as the gated recurrent unit (GRU) [37], the structure of which is shown in Figure 5.

Similar to long short-term memory (LSTM), each GRU unit contains an update gate, reset gate, and candidate hidden state, effectively capturing the long-term dependencies in sequences. Specifically, assuming that there are h hidden units at a given time step t and for a mini-batch input $X_t\in {\mathbb{R}}^{n\times d}$ (where n is the number of samples, and d is the number of input features), as well as the hidden state from the previous time steps $H_{t-1}\in {\mathbb{R}}^{n\times d}$, the reset gate $R_t$ and update gate $Z_t$ are calculated as follows:

$$R_t=sigmoid\left(X_t{W}_{xr}+H_{t-1}{W}_{hr}+b_r\right)$$

(2)

$$Z_t=sigmoid\left(X_t{W}_{xz}+H_{t-1}{W}_{hz}+b_z\right)$$

(3)

The Sigmoid function is an activation function that transforms the values of elements to a range of 0 to 1. Therefore, for each element, the values of the reset gate $R_t$ and the update gate $Z_t$ are within [0, 1]. Next, the gated recurrent unit computes a candidate hidden state to aid in the subsequent hidden state computation. This is achieved by performing element-wise multiplication (denoted as ⊙) between the output of the reset gate at the current time step and the hidden state from the previous time step. Specifically, the formula for computing the candidate hidden state ${\tilde{H}}_t\in {\mathbb{R}}^{n\times d}$ at time step t is

$${\tilde{H}}_t=tanh\left(X_t{W}_{xh}+(R_t\odot H_{t-1})W_{hh}+b_h\right)$$

(4)

In this network, each GRU layer’s task is to process a portion of the input data and generate the corresponding hidden states. These hidden states encapsulate abstract features relevant to the sequential data, which can then be used in further operations. Previous research has shown that in addition to CNNs, RNNs excel in handling spectral information in hyperspectral data. Particularly effective in scenarios where data exhibit sequence-like characteristics, RNNs are well-suited for processing spectral information in hyperspectral images as sequential data. Mou et al. [38] previously applied RNNs in hyperspectral image classification tasks, highlighting their significant potential in such classification tasks. The GRU specifically designed for use in this study handles longer sequences and mitigates the vanishing gradient problem. Compared to the three-gate mechanism (forget gate, input gate, and output gate) of LSTM, the GRU has the advantage of being more concise and easier to train when dealing with datasets such as that in this study, which contains 288 spectral bands.

2.2.2. Proposed Method

In this paper, one of the primary reasons for introducing deep learning algorithms for rock classification is the inherent limitations of human-based judgment based on the 28 classes of rock spectral curves. Although rocks are typically classified based on the similarities of their spectral curve trends, different types of rocks may have similar compositions and spectral features. Therefore, relying solely on spectral data in the classification can easily lead to confusion and low classification accuracy.

To this end, this study used HySpex hyperspectral images as research data, combined with rock sample spectra measured by the ASD spectroradiometer, to verify the spectral curves extracted from the corresponding samples in the images. Due to the potential variations in light intensity caused by light source non-uniformity and ambient light during the ASD spectroradiometer data collection process, a white reference was used to preprocess the measurement data, ultimately forming the rock reflectance spectral curves shown in Figure 6 and Figure 7.

As shown in Figure 6a, in this paper, we compare fine-grained granite belonging to category r10 and granite gneiss belonging to category r16 in the initial rock classification system. These two types of rocks have similar compositions but different proportions. However, their spectral trend features are largely consistent, indicating that the similarity of their spectral curves may not be sufficient to distinguish between them. Conversely, Figure 6b compares the andesite belonging to category r15 and the diorite belonging to category r20 in the initial rock classification system. These two types of rock have similar compositions, but their spectral curves are significantly different, illustrating a case of similar composition but different spectra. This further underscores the challenge of relying solely on the spectral features for rock classification.

In addition to the factors mentioned above, another significant influence in pixel-based rock classification models is the variability in the spectral features across pixels on the surface of a rock composed of different materials. Taking potassium feldspar granite (rock type r11) as an example, it is composed of 66% potassium feldspar, 25% quartz, 4% plagioclase, and 5% biotite. As shown in Figure 7, by comparing the scanning results of the HySpex imaging spectrometer with the measurements obtained using an Analytical Spectral Devices, Inc. (ASD) spectrometer, it was found that different locations on the same rock had different spectral features. Specifically, the spectral feature curves of pixel 1 (Figure 7a), pixel 2 (Figure 7b), and pixel 3 (Figure 7c) exhibit significant differences, which also differ from the spectral curve of the entire rock (Figure 7d).

This phenomenon reveals that in the analysis of high-spectral rock images, pixels containing different proportions of the components of the rock may exhibit varying spectral responses. This variability can potentially impact the accuracy of pixel-based rock classification models. Therefore, in the design and training of these models, it is crucial to consider how to handle these pixel-level spectral differences to enhance the accuracy and robustness of the rock classification.

To address the aforementioned issues in existing hyperspectral rock classification research, where relying solely on spectral information leads to the ‘same material with different spectra’ and ‘different materials with the same spectra’ problems, as well as the regional spectral differences caused by rocks being mineral composites, this study proposes a network model capable of extracting both spectral and spatial information from rocks, integrating and learning from these two types of features.

The proposed network model adopts a dual-branch structure in which the CNN and RNN serve as spatial and spectral feature extractors, respectively. These branches are subsequently connected through fully connected layers for the classification task. The specific structural design of the model is illustrated in Figure 8. Regarding the data input, these dual-branch networks require two types of data. One involves obtaining raw hyperspectral image data, performing principal component analysis for dimensionality reduction, and segmenting it into uniformly sized three-dimensional cubic data as the input for spatial feature extraction by the CNN. The other type involves obtaining segmented raw hyperspectral data, extracting central pixel blocks, and subjecting them to appropriate dimensionality reduction. These processed data serve as spectral features and are input into the RNN.

In the architecture design of this network model, the upper branch is primarily composed of a CNN. This branch includes a series of 2-D convolutional layers tailored for spatial dimension convolution operations on the input data. To ensure these convolutional layers effectively extract spatial features, a 2-D CNN spatial feature extraction branch based on PCA dimensionality reduction was designed. This branch focuses on extracting the spatial details related to rock edges, textures, and positions. By applying PCA to reduce the dimensionality of the input data in the spectral dimension, the model not only reduces computational complexity but also removes redundant information, thereby enhancing model performance. These convolutional layers vary in parameters and kernel sizes and include batch normalization and activation functions. Notably, a significant departure from traditional 2-D convolutional neural networks is the omission of pooling layers, allowing for a higher spatial resolution to be maintained during feature propagation between the convolutional layers. In dense prediction tasks, this approach not only reduces computational complexity but also proves advantageous in preserving more detailed information by avoiding spatial dimension reduction through pooling.

The lower branch is primarily composed of an RNN, which specifically consists of gated recurrent unit layers. When spectral data are fed into this branch, they are initially processed by a 1-D CNN at the front end. The role of this 1-D CNN is to extract the central pixel blocks and reduce the dimensionality of the input data blocks. When operating on spectral dimension data, the 1-D CNN aims to capture the spectral features and patterns from the input data. Considering that recurrent neural networks may encounter challenges such as gradient vanishing or exploding, especially when dealing with extended sequences, their ability to capture long-range dependencies is limited. The core of this network structure consists of three independent GRU layers, collectively forming a Tri-GRU, which is crucial for handling the temporal sequence features of the input data. In this context, the spectral dimension features of the hyperspectral images can be analogized to temporal sequence features. Each GRU layer includes a unidirectional GRU unit that is adept at capturing sequential patterns in the input data. By combining and stacking GRU units across different layers, the network effectively learns the patterns and features in the input data at different time scales. Recurrent neural networks have been proven to be particularly advantageous in the sequence modeling and processing of time series data, aiding the model’s understanding of temporal relationships within the data.

The pixel-by-pixel classification model, composed of the two branches described above, effectively utilizes and integrates both spatial and spectral features of hyperspectral rock data. Each rock’s spatial features (such as the shape, size, and surface texture) are inconsistent. The model avoids heavily relying on these specific spatial features during training while leveraging the spatial information from neighboring pixels to correct the misclassifications of spectral information by the GRU on individual pixels. Ultimately, the fusion strategy of feature concatenation under the complementary nature of these two types of information results in a more consistent and accurate overall classification.

3. Experiments and Discussion

3.1. Experiment

3.1.1. Experimental Environment and Evaluation Metric

The experiments in this study were conducted on a platform consisting of an Intel(R) Xeon(R) Platinum 8255C CPU, an NVIDIA GeForce RTX 2080Ti GPU, and the Ubuntu 20.04 operating system. All experiments were performed using PyTorch 1.11.0, Python 3.8, and CUDA 11.3. The specific Python modules and libraries used include numpy 0.2.4, scikit-learn 1.3.2, and torch 1.11.0+cu113.

The focus of this study was the classification of hyperspectral images. Therefore, common metrics in this field, including the overall accuracy (OA), average accuracy (AA), and kappa coefficient, were employed to evaluate the quality of the classification results. The OA represents the proportion of the correctly classified samples to the total number of samples, and it is calculated as follows:

$$OA={\displaystyle \frac{\sum _{i=1}^n{h}_{ii}}{\sum _{i=1}^n{N}_i}}$$

(5)

where n is the number of rock categories in the image, $N_i$ is the number of pixels in the ith rock category, and $h_{ii}$ is the number of pixels correctly classified in the ith category. The AA is the sum of the proportions of the samples in each category relative to the total number of samples divided by the number of categories. It is calculated as follows:

$$AA={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\displaystyle \frac{h_{ii}}{N_i}}$$

(6)

The kappa coefficient is an indicator of the consistency of the model and is used to measure the effectiveness of the classification. It is calculated as follows:

$$Kappa={\displaystyle \frac{N\sum _{i=1}^r{x}_{ii}-\sum _{i=1}^r\left(x_{i+}{x}_{+i}\right)}{N^2-\sum _{i=1}^r\left(x_{i+}{x}_{+i}\right)}}$$

(7)

where N is the total number of samples, r is the number of rows in the confusion matrix, $x_{i+}$ and $x_{+i}$ are the sums of the elements in the ith row and ith column of the confusion matrix, respectively, and $x_{ii}$ is the value in the ith row and ith column of the confusion matrix.

3.1.2. Selection of the Network Backbone

Three main network architectures can be chosen: the 2-D CNN, GRU, or 2-D CNN-GRU fusion network. Compared to the 2-D CNN, the GRU has advantages in the spectral dimension of the hyperspectral data, but it cannot compensate for the spatial information that is ignored in some specific datasets. The 2-D CNN-GRU fusion network considers both the overall classification performance and the efficiency.

Table 3. Comparison of three network structures.

Model	OA (%)	AA (%)	Kappa × 100
2-D CNN	84.011 ± 6.751	82.615 ± 7.349	83.30 ± 7.1
GRU	84.610 ± 2.253	83.522 ± 2.666	83.90 ± 2.4
2-D CNN-GRU	97.925 ± 2.831	97.956 ± 2.915	97.80 ± 3.0

presents the classification performances of these three network structures on the hyperspectral rock dataset.

Table 3 shows that both the OA and AA were higher when trained using the 2-D CNN-GRU backbone network compared to training using the 2-D CNN or GRU individually. This improvement is attributed to their respective strengths in handling different data characteristics, and using them in combination compensates for the shortcomings of each and enhances the overall performance. The 2-D CNN excels in extracting spatial features from hyperspectral images by capturing the relationships and texture information between the pixels, while the GRU captures the temporal information between bands to better understand the spectral evolution.

3.1.3. Parameter Settings

The model weights were initialized using He normal distribution, with the biases set to zero. For layers that do not include biases (such as the batch normalization layers), bias initialization was not necessary. The optimizer used was Adam, with an initial learning rate of 0.001, which was dynamically adjusted. During model training, L2 regularization, early stopping, and cross-validation techniques were employed to prevent overfitting. Specifically, L2 regularization was implemented by setting the weight decay parameter of the optimizer to 0.01. Early stopping involved monitoring the accuracy of the validation set during training, and training was halted when the accuracy ceased to improve over several epochs. In this study, the optimal number of training epochs was determined to be 50.

For dataset processing, to make more effective use of the data, k-fold cross-validation was employed to randomly split the training and testing image samples, thereby validating the model’s performance across multiple subsets. Specifically, 5-fold cross-validation was applied to split the dataset across the entire hyperspectral image. In each fold, 20% of the pixels were randomly selected as the training set for model training, and the remaining pixels were used as the testing set. This process is repeated for all five folds. Additionally, within the testing set, 10% of the samples were selected as a validation set for parameter optimization.

One branch of the convolutional network in the model received fixed-size hyperspectral spatial neighborhoods as the input data. The amount of data the model received depended on the size of these patches, which may significantly impact the final classification performance. Therefore, in this study, experiments were conducted on the dataset with different batch sizes and three patch sizes (5, 7, and 9) to determine the optimal parameter size of the model. The results are shown in

Table 4. OA (%) of training datasets with different adjacent pixel block sizes and batch sizes.

Neighboring Pixel Block Size	Batch Size
	64	128	256
5 × 5	96.451	95.328	95.986
7 × 7	96.894	97.543	97.325
9 × 9	97.145	97.249	96.874

.

As shown in Table 4, the model achieved the highest classification accuracy when the dataset batch size was 128 and the adjacent pixel block size was 7. Therefore, in the subsequent experiment, all of the models were configured using these dataset size parameters.

3.1.4. Analysis of the Impact of PCA-Preserved Components

In this study, PCA dimensionality reduction was applied before feeding the data into the spatial feature extraction branch of the model. PCA significantly reduces the number of spectral features while preserving the primary spatial features in the hyperspectral data. Since the convolution operations of the 2-D CNN are influenced by the spectral dimension of the input pixel blocks, this could impact the computational complexity and classification performance. Therefore, based on the aforementioned parameters, experiments were conducted with different dimensions of input data for this branch (4, 16, 32, and the original 288 dimensions without reduction) to determine the optimal number of components to retain after PCA dimensionality reduction.

As shown in

Table 5. The influence of different preserved components on classification accuracy and training time after PCA dimensionality reduction.

Retained Components	OA (%)	AA (%)	Kappa × 100	Average Training Duration (s)
4	89.802 ± 4.173	88.043 ± 5.922	89.00 ± 5.3	1698.775
16	97.925 ± 2.831	97.956 ± 2.915	97.80 ± 3.0	1764.083
32	93.119 ± 3.598	92.264 ± 6.744	92.60 ± 4.4	1852.693
288	92.563 ± 2.970	93.295 ± 2.642	92.20 ± 3.1	2023.253

, retaining too few or too many components after PCA dimensionality reduction adversely affects the model’s final classification performance, and the training time of the model decreases as the number of retained components is reduced. This indicates that the convolution operations of the 2-D CNN are influenced by the spectral dimension of the input pixel blocks, requiring a trade-off between computational complexity and classification accuracy. Therefore, in this study, the spectral dimension of the training data input into this branch was ultimately reduced to 16, achieving the optimal balance.

3.2. Results and Discussion

To evaluate the effectiveness of the rock classification algorithm developed in this study, we compared the developed model with existing models that are capable of classifying hyperspectral data, including Resnet-18 [39], 3-D CNN [40], Hamida [41], HybridSN [42], and CAE-SVM [43]. As previously described, we used a 5-fold cross-validation method to partition the data required for the training, validation, and testing of each model. In each fold, 20% of the data were allocated for training, and the remaining 80% of the data were randomly sampled to form a validation set, which included 10% of the data, leaving 70% of the data for testing. This experiment was repeated five times, and the results of all of the runs were averaged. The results are presented in

Table 6. Classification results of the different methods on the rock datasets.

Method	OA (%)	AA (%)	Kappa × 100
Resnet-18	87.995 ± 7.486	85.678 ± 8.526	87.40 ± 7.8
3-D CNN	94.075 ± 3.286	92.505 ± 3.717	93.80 ± 3.4
Hamida	88.441 ± 6.411	88.003 ± 6.324	87.90 ± 6.7
HybridSN	90.010 ± 2.212	89.078 ± 2.469	89.50 ± 2.3
CAE-SVM	91.969 ± 2.798	92.228 ± 2.065	91.60 ± 2.9
Ours	97.925 ± 2.831	97.956 ± 2.915	97.80 ± 3.0

.

As can be seen from Table 6, the 2-D CNN-GRU model outperforms the other models in terms of the OA and AA. Additionally, its kappa coefficient exceeds those of the other models, demonstrating the suitability and superiority of this model in the context of hyperspectral rock datasets.

Figure 9 shows representative classification prediction maps from one fold of the models, with the classification accuracies being close to the average. The classification maps of the rock dataset display varying accuracies for the different rock types among the different models. Each model has specific strengths and weaknesses in its ability to classify the rock types. Notably, the classification maps produced using the 3-D CNN and the model developed in this study are smoother compared to those obtained using the other models, and they contain fewer misclassified pixels. The other four models exhibit noticeable pixel misclassifications due to their insufficient feature extraction capabilities.

Figure 10 presents the results of the comparison of the overall accuracies (OAs) of each classification model for different training set proportions. In Figure 10, the vertical axis represents the OA, and the horizontal axis represents the proportion of the training sample. It can be seen from Figure 10 that the OAs of the model developed in this study are higher than those of the other methods for five different proportions: 1%, 5%, 10%, 15%, and 20%. This indicates that our model is superior for both small and large sample classifications.

Table 7. The classification results of each type of rock on the rock dataset using different methods.

Class Name	Resnet-18	3-D CNN	Hamida	HybridSN	CAE-SVM	Ours
r1	89.2 ± 10.4	97.6 ± 3.4	93.0 ± 4.5	96.0 ± 4.3	94.1 ± 3.8	98.6 ± 1.9
r2	88.8 ± 5.7	94.2 ± 5.9	93.1 ± 2.3	94.9 ± 3.3	95.1 ± 1.6	98.8 ± 1.6
r3	99.9 ± 0.1	99.0 ± 1.1	99.2 ± 1.5	99.9 ± 0.2	95.9 ± 4.7	83.8 ± 32.4
r4	99.8 ± 0.2	99.3 ± 0.5	89.3 ± 21.2	99.8 ± 0.2	90.9 ± 10.4	94.1 ± 11.5
r5	92.1 ± 5.4	95.9 ± 2.3	81.0 ± 13.9	89.0 ± 7.9	76.3 ± 15.3	99.2 ± 0.6
r6	98.6 ± 1.4	98.7 ± 0.6	98.8 ± 1.8	97.3 ± 1.5	97.0 ± 1.6	99.5 ± 0.9
r7	92.7 ± 2.6	92.2 ± 1.7	94.4 ± 2.7	95.2 ± 1.2	90.8 ± 5.0	99.4 ± 0.1
r8	98.4 ± 1.3	97.2 ± 2.3	98.5 ± 1.9	98.9 ± 0.4	97.1 ± 1.5	99.9 ± 0.1
r9	99.9 ± 0.1	99.0 ± 1.2	99.8 ± 0.2	99.4 ± 0.5	97.7 ± 2.0	99.4 ± 1.2
r10	92.9 ± 2.1	96.1 ± 4.1	84.9 ± 11.4	94.9 ± 2.8	94.4 ± 2.9	99.5 ± 0.4
r11	96.1 ± 2.9	98.4 ± 1.1	65.9 ± 37.4	94.1 ± 3.5	94.0 ± 3.0	99.8 ± 0.1
r12	98.8 ± 0.6	97.4 ± 1.8	97.8 ± 2.2	94.2 ± 1.5	95.6 ± 4.3	99.8 ± 0.3
r13	85.5 ± 3.4	88.1 ± 13.9	87.1 ± 5.3	80.9 ± 5.4	83.2 ± 7.1	99.5 ± 0.1
r14	89.4 ± 6.1	97.5 ± 2.8	95.3 ± 5.5	88.9 ± 8.1	90.3 ± 8.1	97.5 ± 4.8
r15	98.2 ± 2.2	98.2 ± 1.7	97.5 ± 2.0	98.1 ± 1.8	96.5 ± 3.0	99.5 ± 0.8
r16	87.3 ± 5.2	98.3 ± 0.8	84.7 ± 7.4	89.0 ± 4.2	93.1 ± 2.2	99.3 ± 0.6
r17	78.4 ± 20.0	94.9 ± 3.3	82.8 ± 9.0	93.1 ± 2.7	93.8 ± 3.8	96.7 ± 5.5
r18	98.0 ± 0.5	95.6 ± 5.7	99.6 ± 0.2	99.8 ± 0.3	94.5 ± 3.7	100.0 ± 0.0
r19	96.1 ± 1.0	94.9 ± 3.4	98.4 ± 0.6	96.7 ± 4.8	97.6 ± 1.8	97.9 ± 3.7
r20	75.2 ± 12.1	96.5 ± 2.3	76.0 ± 34.6	92.0 ± 4.8	94.0 ± 2.4	99.4 ± 0.8
r21	78.7 ± 17.1	96.4 ± 1.1	85.1 ± 9.2	78.0 ± 12.9	86.1 ± 8.0	99.5 ± 0.6
r22	78.9 ± 11.8	88.2 ± 4.0	80.2 ± 20.5	89.9 ± 5.2	90.4 ± 2.3	99.1 ± 0.5
r23	99.6 ± 0.2	98.8 ± 1.5	99.0 ± 0.9	96.3 ± 1.2	98.6 ± 0.6	99.9 ± 0.1
r24	88.8 ± 2.5	93.7 ± 2.8	76.6 ± 13.2	90.6 ± 2.2	86.7 ± 6.8	93.9 ± 11.8
r25	67.5 ± 10.3	91.1 ± 10.3	77.0 ± 14.5	76.0 ± 9.7	91.6 ± 7.0	99.4 ± 0.5
r26	76.3 ± 9.0	90.9 ± 3.0	83.5 ± 6.9	73.3 ± 9.6	88.3 ± 7.1	97.6 ± 2.4
r27	52.7 ± 10.1	87.3 ± 8.0	74.2 ± 12.3	73.5 ± 9.4	74.9 ± 19.7	91.9 ± 12.3
r28	56.8 ± 6.2	53.5 ± 22.1	66.8 ± 15.1	45.4 ± 11.3	78.3 ± 9.8	88.1 ± 13.4

presents more detailed classification results for each rock type. As can be seen from the data presented in Table 7, all six models have higher error ranges and lower classification accuracies in identifying the rock types r27 and r28. This could be due to the poor spectral matching characteristics within each block for these two rock types, resulting in varying accuracies across the five-fold dataset tests. Similarly, rock types r21 and r22 have lower classification accuracies for each model’s classification experiments. Additionally, in one fold of the dataset, the model developed in this study has a lower accuracy in identifying rock type r3, which may be due to the poor representation in the sample selection during the operation of the random sampling algorithms, resulting in significant accuracy errors in the final assessment. The other five models have higher accuracies in identifying rock type r3. However, the model developed in this study has an excellent classification performance for most of the rock classifications. For example, for rock type r5, which contains only one rock sample, the other five models have lower accuracies due to difficulties in handling the categories containing few rock samples, whereas the model developed in this study has significant advantages and higher classification accuracy in handling such categories. In contrast, rock type r26 includes six rock samples with highly similar spectral curves, yet the model developed in this study still achieves a high classification accuracy. This demonstrates that the model developed in this study effectively classifies both sample categories containing several rock samples and categories with few rock samples.

4. Conclusions

This study addresses the typical issues of ‘same spectrum, different materials’ and ‘different materials, same spectrum’ by proposing a dual-branch multidimensional feature network model that combines CNN and RNN based on the theory of integrating spatial and spectral information from hyperspectral image data. The model extracts spatial features of rocks through a 2-D CNN branch with a PCA dimensionality reduction module, while another branch, primarily using a GRU network, extracts spectral features of the rocks. The features extracted from both branches are then fused and classified, effectively solving the issue of insufficient feature utilization when relying solely on the spectral features for rock classification. Furthermore, this hyperspectral rock classification research based on laboratory platform data further validates the effectiveness of the preliminary rock classification system, achieving an overall classification accuracy of 97.925% and an average classification accuracy of 97.956% on a dataset containing 81 types of igneous and metamorphic rocks, significantly outperforming other classification models.