Hyperspectral rock classification method based on deep ensemble learning algorithm

Abstract

The invention discloses a hyperspectral rock classification method based on a deep ensemble learning algorithm. The method comprises the following steps of: 1, acquiring a hyperspectral data set of 81 types of rock samples from a hyperspectral rock standard database, performing dimension reduction by applying principal component analysis, and segmenting the hyperspectral data set into a three-dimensional cube as spatial features to be input into a 2D convolutional neural network; meanwhile, the center pixel block subjected to dimension reduction processing serves as a spectral feature to be input into a gating circulation unit; 2, connecting a 2D convolutional neural network and a gating circulation unit in series, introducing a full connection layer to fuse spatial features and spectral features, and optimizing model performance in combination with an AdaBoost algorithm; and 3, dividing the hyperspectral data set subjected to principal component analysis dimension reduction into a training set and a test set, and training the hyperspectral rock classification model in batches. According to the method, the space and spectral characteristics of the rock sample are effectively fused, the multi-dimensional information of the hyperspectral rock image is fully utilized, the phenomena of same object and different spectrum and same spectrum and foreign matter are reduced, and the classification accuracy is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of hyperspectral image processing and deep learning technology, and specifically relates to a hyperspectral rock classification method combined with a deep ensemble learning algorithm, which is suitable for intelligent rock classification in geological exploration, mineral resource assessment and tectonic evolution research. Background Art

[0002] As the fundamental building blocks of Earth's surface materials, accurate rock classification is crucial for fields such as mineral resource assessment, geological tectonic evolution research, and environmental monitoring. Traditional rock classification methods rely heavily on field macroscopic observations and laboratory microscopic identification, such as spectral matching and polarizing microscopy. However, these methods suffer from low efficiency, strong subjectivity, and difficulty adapting to complex geological environments. These methods exhibit significant limitations, particularly when applied to large-scale geological mapping or high-precision resource exploration.

[0003] In recent years, hyperspectral imaging technology, with its nanometer-scale spectral resolution (λ / 100) and wide spectral coverage, has been able to simultaneously analyze both microscopic compositional differences and macroscopic spatial structural characteristics of rocks, providing a powerful technical means for intelligent rock classification. This technology has been widely used in geological mapping, mineral resource exploration, and other fields. However, as research deepens, existing methods have gradually exposed three core challenges:

[0004] (1) Weak data foundation: The existing hyperspectral rock dataset has limited coverage, making it difficult to fully reflect the spectral variability and secondary alteration effects of igneous and metamorphic rock subtypes, and the ability to distinguish highly similar rock types is insufficient.

[0005] (2) Generalization problem of small samples: In actual geological scenarios, due to sampling condition limitations, hyperspectral rock images often face the problem of insufficient sample size, which makes the model prone to overfitting and limited classification accuracy.

[0006] (3) Spectral confusion: When the number of rock types exceeds 50, the differences in spectral characteristics are subtle, the phenomena of "same spectrum, different objects" and "same object, different spectrum" increase significantly, and the classification accuracy of mainstream models generally decreases. For example, in ophiolite suites, peridotite and gabbro show extremely high similarity in specific bands; while in granite, potassium feldspar shows different spectral characteristics due to differences in crystallization conditions. Summary of the Invention

[0007] In response to the problems existing in the prior art, the present invention proposes a hyperspectral rock classification method based on a deep ensemble learning algorithm, comprising: first, obtaining a hyperspectral dataset of 81 types of rock samples from a hyperspectral rock standard database, using principal component analysis to reduce the dimensionality, and then dividing it into three-dimensional cubes as spatial features and inputting them into a 2D convolutional neural network; at the same time, the central pixel block after dimensionality reduction is input into a gated recurrent unit as a spectral feature; then, the 2D convolutional neural network and the gated recurrent unit are connected in series, and a fully connected layer is introduced to fuse spatial features and spectral features, and the model performance is optimized in combination with the AdaBoost algorithm; subsequently, the hyperspectral dataset after principal component analysis dimensionality reduction is divided into a training set and a test set according to a certain ratio, and the hyperspectral rock classification model is trained in batches, and the learning rate and weak classifier weights are dynamically adjusted during the training process until the model converges; finally, the classification effect of the trained model is evaluated using the test set, and a classification result graph is generated to comprehensively test and demonstrate the classification ability of the model.

[0008] Furthermore, the hyperspectral rock classification model includes a spatial feature extraction module, a spectral sequence feature extraction module, and an AdaBoost optimization module.

[0009] The spatial feature extraction module: This module constructs multiple customized 2D convolution layers by streamlining the number of convolution channels and layers. The convolution layers use different parameters and kernel sizes, and add batch normalization layers and ReLU activation functions after each convolution layer. At the same time, the pooling layer is omitted to maintain a high spatial resolution. Different 2D convolution layers efficiently capture multi-scale spatial structural features from the original hyperspectral image data after PCA dimensionality reduction processing to generate spatial features. Figure I ; These spatial features are then Figure I Fusion is performed to obtain spatial features that fuse multi-scale spatial features Figure II , in preparation for the subsequent fusion with spectral features; when fused with spectral features, the spatial features Figure II Enter the fully connected layer.

[0010] The spectral sequence feature extraction module is composed of 1D-CNN and GRU. 1D-CNN processes spectral data with a specific convolution kernel and adjusts the step size and padding parameters. After batch normalization and ReLU activation, preliminary spectral features are generated. Figure I The GRU layer uses the reset gate and update gate mechanism to stack multiple layers to capture spectral features at different time scales and generate spectral features with rich spectral response patterns. Figure II ; Then the spectral characteristics Figure II After batch normalization and ReLU activation function integration optimization, the spectral features are obtained Figure III ; Finally, the spectral features Figure III Input fully connected layer, and spatial features of 2D-CNN branch Figure II Deep fusion, the neurons in the fully connected layer are connected to the two types of feature map elements to obtain a fused feature map, which improves the classification performance of hyperspectral rock images.

[0011] The AdaBoost optimization module plays a role in dynamically adjusting sample weights and model decision weights in the hyperspectral rock classification model. In each iteration, AdaBoost adjusts sample weights based on the current classification results based on the fully connected fusion feature map, thereby increasing the weights of misclassified samples. At the same time, AdaBoost enables the model to better cope with noise and outliers when processing the fusion feature map. For samples with "same spectrum but different objects" or "same object but different spectrum", the model focuses on relevant features, thereby improving model robustness. In addition, the AdaBoost optimization module calculates and adjusts the weights of weak classifiers. In the process of weighted linear superposition and integration of weak classifiers, it highlights the role of important spectral features in the fusion feature map, suppresses interference, and improves classification accuracy and reliability.

[0012] Furthermore, the GRU, as a variant of LSTM, has a simpler gating mechanism, which only contains update gates and reset gates; at a given time step t, a small batch input and the hidden states of the previous time steps In case of a fault, reset gate R t and update gate Z t The calculation formula is as follows:

[0013] R t =sigmoid(X t W xr +H t-1 W hr +b r ) (1)

[0014] Z t =sigmoid(X t W xz +H t-1 W hz +b z ) (2)

[0015] Where n is the number of samples, d is the number of inputs, and h is the hidden unit dimension.

[0016] In formula (1), W xr represents the weight matrix from the input layer to the reset gate; W hr represents the recursive weight from the hidden state to the reset gate; b r is the bias vector for the reset gate.

[0017] In formula (2), W xz Represents the projection weight input to the update gate; Whz represents the recursive weight from the hidden state to the update gate; b z Bias for the update gate.

[0018] The Sigmoid function is used as an activation function to map values between 0 and 1, controlling the gate mechanism in the GRU; this makes the reset gate R t and update gate Z t The output value is in the range [0,1]; GRU calculates the candidate hidden state To assist in updating the hidden state, this operation is performed by performing element-wise multiplication between the output of the reset gate at the current time step and the hidden state of the previous time step; specifically, the candidate hidden state at time step t The calculation formula is as follows:

[0019]

[0020] In formula (3), W xh Represents the projection weight of the input to the candidate state; W hh represents the recursive weight of the hidden state after reset; b h is the candidate state bias, and ⊙ represents element-wise multiplication.

[0021] Furthermore, the AdaBoost optimization module implements a dual enhancement mechanism by dynamically adjusting sample weights and model decision weights. By iteratively adjusting the sample weight distribution, the model is forced to continuously focus on difficult samples. The specific execution process is as follows:

[0022] First, initialize the sample weights; assuming that the dataset D contains M samples, the initial weight of each sample is equal to That is, the weight of the i-th sample in the first round of iteration.

[0023] Then use the current sample weight to train a weak classifier c t (x), and calculate the weighted error rate r of the weak classifier on the training set t , formula (4) is as follows:

[0024]

[0025] In formula (4), I(·) is an indicator function, which takes the value of 1 when the prediction is wrong and 0 otherwise; i represents the eigenvector of the i-th sample, so c t (x i ) is the weak classifier for sample c in the tth round i The predicted value, y i That is, the input data represents the label of the i-th sample, that is, the output target that the model needs to predict.

[0026] According to the weighted error rate r t Calculate the weight a of the weak classifier t , as shown in formula (5):

[0027]

[0028] Weight a t Reflects the importance of the weak classifier in the final decision, the weighted error rate r t The lower it is, the higher the weight.

[0029] Then, update the sample weights so that the weights of misclassified samples increase, as shown in formula (6):

[0030]

[0031] In each iteration of the AdaBoost algorithm, the sample weights need to be normalized after being updated, that is, each sample weight value is divided by the current total weight so that the sum of the normalized weights is constant to 1; when all the preset iteration rounds are completed, the algorithm will calculate the corresponding weight a of each weak classifier according to its error rate during training. t , and finally all weak classifiers are integrated into a strong classifier by weighted linear superposition; its mathematical expression is shown in formula (7):

[0032]

[0033] Where sign(·) is a sign function used to determine the final classification result; T represents the total number of iterations, that is, the total number of weak classifiers used in the training process.

[0034] Furthermore, the batch training model is performed until convergence and stability are achieved, specifically including:

[0035] In terms of model parameter setting and training strategy, the AdaBoost algorithm uses 10 decision tree stumps as base classifiers. Using three classifiers has been empirically determined to achieve the optimal balance between avoiding overfitting and reducing computational overhead. The C-RNN model uses the He normal distribution to initialize the convolutional layer weights to ensure stable gradient propagation during the initial training phase. Layers containing bias terms are initialized to zero, while layers without bias terms are not initialized. Training is capped at 50 epochs, balancing classification performance and computational efficiency. The Adam algorithm is used for optimization, with an initial learning rate of 0.001 and dynamic adjustment. Data is partitioned using 5-fold cross-validation, with 20% of pixels randomly selected for training, 70% for testing, and 10% for validation at each iteration. These carefully designed parameters and strategies further enhance the performance and reliability of the hyperspectral rock classification method based on deep ensemble learning.

[0036] Finally, the test set is used to evaluate the classification effect and generate the classification results.

[0037] The beneficial effects of the present invention are as follows:

[0038] (1) Through the combination of multiple models, the multidimensional information of 81 rock classification categories is fully utilized, the spatial and spectral characteristics of rock samples are effectively integrated, and the risk of overfitting caused by the reduction in the number of rock classification samples and the occurrence of "same spectrum, different objects" and "same object, different spectrum" phenomena are reduced;

[0039] (2) The introduction of the AdaBoost algorithm enhances the stability and generalization ability of the model, especially when the number of samples is small and the categories are numerous. It can effectively reduce overfitting, improve the model's discrimination ability, and significantly improve the recognition accuracy and robustness of small sample categories.

[0040] (3) A lightweight GRU structure is proposed, which reduces the number of parameters by about 40%, significantly reducing computing resources and time consumption while maintaining high classification accuracy.

[0041] Experimental results show that the overall accuracy (OA) of the proposed method reaches 92.55%, the average accuracy (AA) reaches 92.25%, and the Kappa coefficient reaches 92.50% in 81 types of rock classification tasks, which is significantly better than the existing mainstream models. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions used in the present invention, the following briefly introduces the drawings required for the technical description.

[0043] Figure 1 It is a schematic diagram of the network flow designed by the embodiment of the present invention;

[0044] Figure 2 It is a schematic diagram of the network structure of the present invention;

[0045] Figure 3 Schematic diagram of the spatial feature extraction module used in the present invention;

[0046] Figure 4 Schematic diagram of the spectrum sequence feature extraction module used in the present invention;

[0047] Figure 5 It is a schematic diagram of the AdaBoost algorithm of the present invention;

[0048] Figure 6 This is a diagram of the classification effects of the present invention and other models under different RNN branches on 81 types of rock data sets.

[0049] Figure 7 This is a diagram showing the classification effects of the present invention and other different models on 81 types of rock data sets.

[0050] in, Figure 6 -a, Figure 6 -b, Figure 6 -c, Figure 6 -d are 2D-BRNN-AdaBoost, 2D-LSTM-AdaBoost, 2D-1D CNN-AdaBoost and the model CRNN-AdaBoost in this paper.

[0051] in, Figure 7 -a, Figure 7 -b, Figure 7 -c, Figure 7 -d, Figure 7 -e, Figure 7 -f are 3D-CNN, Bi-CLSTMSSRN, DenseNet, the model CRNN-AdaBoost in this paper and the real labels of 81 types of rocks. DETAILED DESCRIPTION

[0052] The present invention will be further described below with reference to the accompanying drawings and examples.

[0053] This paper proposes a hyperspectral rock classification model based on deep ensemble learning. This model fuses spatial and spectral features through a dual-branch network structure and combines it with the AdaBoost algorithm to optimize classification performance. The model comprises three core modules: a spatial feature extraction module employs a multi-scale 2D-CNN to capture rock texture and edge information; a spectral sequence feature module extracts temporal spectral features through a combination of a 1D-CNN and a GRU; and an AdaBoost module dynamically adjusts sample weights and classifier decision weights to enhance robustness against complex samples such as "same spectrum, different objects." Data preparation and input utilize PCA dimensionality reduction to generate spatial cube inputs and center pixel spectral sequence inputs. Training incorporates 5-fold cross-validation, dynamic learning rate adjustment, and a He normal distribution initialization strategy. Finally, classification accuracy is improved through a weighted ensemble of weak classifiers. Through feature fusion and a dual weight optimization mechanism, this model effectively addresses the challenge of synergistically utilizing spatial and spectral features in hyperspectral rock images, while also effectively mitigating the "same spectrum, different objects" phenomenon. Ablation experiments with deformable convolutional layers confirm that the adaptive sample weighting mechanism significantly improves recognition accuracy and robustness for small sample sizes.

[0054] like Figure 1 As shown, the specific implementation steps are as follows:

[0055] Step 1: A hyperspectral rock classification method based on deep ensemble learning. The network model requires two types of data:

[0056] The first is the original hyperspectral image data, which is divided into three-dimensional cube data of uniform size after dimensionality reduction through principal component analysis to serve as the spatial features of the input 2D convolutional neural network;

[0057] The second type of data involves extracting central pixel blocks from the segmented raw hyperspectral data and performing appropriate dimensionality reduction on them. These processed data are used as spectral features and input into the gated recurrent unit module.

[0058] Step 2: Connect the 2D convolutional neural network (2D-CNN branch) and the gated recurrent unit (RNN branch) in series, introduce a fully connected layer to fuse spatial and spectral features, and combine the AdaBoost algorithm to optimize performance. Figure 2 As shown in Figure 1, the hyperspectral rock classification model includes a spatial feature extraction module, a spectral sequence feature extraction module, and an AdaBoost optimization module, specifically including:

[0059] like Figure 3 The figure shows the spatial feature extraction module. This module streamlines the number of convolutional channels and layers. It consists of two 2D convolutional layers using 7x7 kernels. Large kernels are beneficial for capturing large-scale spatial features such as rock texture and edges. Batch normalization layers and ReLU activation functions are added after the convolutional layers, while pooling layers are omitted to maintain high spatial resolution.

[0060] Different 2D convolutional layers efficiently capture multi-scale spatial structural features from the original hyperspectral image data after PCA dimensionality reduction and generate spatial features Figure I These spatial features are then Figure I Fusion is performed to obtain spatial features that fuse multi-scale spatial features Figure II , in preparation for subsequent fusion with spectral features. When fused with spectral features, the spatial features Figure II Enter the fully connected layer.

[0061] like Figure 4 The figure shows the spectral sequence feature extraction module, which uses a 1D-CNN and GRU combined architecture to effectively capture spectral sequence features.

[0062] First, the hyperspectral image is preprocessed, and the central pixel block is extracted and input into the 1D-CNN. In this part of the 1D-CNN, a 3x1 convolution kernel is used, and the appropriate step size and padding parameters are set according to the actual situation. At the same time, a specific convolution kernel can be used and the step size and padding parameters can be adjusted to process the spectral data. After stacking multiple convolution layers, the spectral features are extracted and reduced in dimension, redundant information is removed, and the preliminary processed data is output. After batch normalization and ReLU activation, the preliminary spectral features are generated. Figure I .

[0063] Then, GRU uses gated recurrent units to model the temporal dependencies of the spectral sequence and extract multi-level spectral response patterns. As a variant of LSTM, GRU has a simpler gating mechanism that only contains update gates and reset gates.

[0064] (n is the number of samples, d is the number of inputs) and the hidden states of the previous time steps (h is the hidden unit dimension), reset the gate R t and update gate Z t The calculation formula is as follows:

[0065] R t =sigmoid(X t W xr +H t-1 W hr +b r ) (1)

[0066] Z t =sigmoid(X t W xz +H t-1 W hz +b z ) (2)

[0067] In formula (1), W xr (x represents input, r represents reset gate) represents the weight matrix from the input layer to the reset gate; W hr Represents the recursive weight from the hidden state to the reset gate, modeling the band correlation across time steps; b r To reset the bias vector of the gate, control the default threshold for gate activation.

[0068] In formula (2), W xz represents the projection weight input to the update gate to extract diagnostic absorption features; W hz Represents the recursive weight from the hidden state to the update gate, maintaining the memory of key spectral features; b z For updating gate biases, positive initialization can enhance the tendency to retain historical states.

[0069] The Sigmoid function is used as an activation function to map the value between 0 and 1, controlling the gate mechanism in the GRU. This makes the reset gate R t and update gate Z t The output value of is in the range [0,1]. GRU calculates the candidate hidden state To assist in updating the hidden state, we perform this operation by performing an element-wise multiplication (denoted as ⊙) between the output of the reset gate at the current time step and the hidden state at the previous time step. Specifically, the candidate hidden state at time step t is The calculation formula is as follows:

[0070]

[0071] In formula (3), W xh represents the projection weight of the input to the candidate state, capturing local spectral details; W hh represents the recursive weight of the hidden state after reset, modeling the nonlinear interaction between bands; b h Bias the candidate states to compensate for spectral baseline shifts.

[0072] In this way, GRU can dynamically adjust the hidden state according to the input sequence, capturing the temporal sequence and multi-level spectral response patterns of the spectral sequence. Stacking multiple layers of GRU can capture spectral features at different time scales and generate features rich in spectral response patterns. Figure II Afterwards, the spectral characteristics Figure II After batch normalization and ReLU activation function, the spectral features are obtained. Figure III .

[0073] Finally, the spectral features Figure III Input fully connected layer, and spatial features of 2D-CNN branch Figure II Deep fusion, the neurons in the fully connected layer are connected to the two types of feature map elements to obtain a fused feature map, which improves the classification performance of hyperspectral rock images.

[0074] like Figure 5 This is the AdaBoost optimization module, which dynamically adjusts sample weights and model decision weights in the hyperspectral rock classification model. This module uses a dual-boost mechanism to iteratively adjust the sample weight distribution, forcing the model to continuously focus on difficult samples. The details are as follows:

[0075] The algorithm execution process is as follows:

[0076] First, initialize the sample weights. Assume that the dataset D contains M samples, and the initial weight of each sample is equal to That is, the weight of the i-th sample in the first iteration. Then use the current sample weight to train a weak classifier c t (x) (t represents the tth round), and calculate the weighted error rate r of the weak classifier on the training set t , formula (4) is as follows:

[0077]

[0078] In formula (4), I(·) is the indicator function, which takes the value 1 when the prediction is wrong and 0 otherwise. i represents the eigenvector of the i-th sample, so c t (x i ) is the weak classifier for the tth round of sample x i The predicted value of y (±1). i That is, the input data represents the label of the i-th sample, that is, the output target that the model needs to predict.

[0079] Calculate the weight a of the weak classifier based on the error rate calculated above t , as shown in formula (5):

[0080]

[0081] This weight reflects the importance of the weak classifier in the final decision. The lower the error rate, the higher the weight. Then, update the sample weight so that the weight of the misclassified sample increases, as shown in formula (6):

[0082]

[0083] At each iteration, AdaBoost adjusts the sample weights according to the current classification results based on the fused feature map, increasing the weights of misclassified samples. At the same time, it enables the model to better cope with noise and outliers when processing the fused feature map. For complex "same spectrum, different objects" or "same object, different spectrum" samples, it adjusts the focus on relevant features to improve model robustness.

[0084] When all the preset iteration rounds are completed, the algorithm will calculate the corresponding combination coefficient (α t ), and finally all weak classifiers are integrated into a strong classifier by weighted linear superposition. Its mathematical expression is shown in formula (7):

[0085]

[0086] Where sign(·) is the sign function used to determine the final classification result. T represents the total number of iterations, that is, the total number of weak classifiers used in the training process.

[0087] AdaBoost calculates and adjusts the weights of weak classifiers, and highlights the features in the process of weighted linear superposition and integration of weak classifiers. Figure III It can effectively suppress interference and improve classification accuracy and reliability by optimizing the important spectral features in the classification.

[0088] Step 3: Use a 5-fold cross-validation strategy to divide the training set, test set, and validation set. Then train the model in batches, dynamically adjust the learning rate and weak classifier weights until the model converges and stabilizes.

[0089] The AdaBoost algorithm configuration uses 10 decision tree stumps as base classifiers, and it is determined that the best balance is achieved when using 3 classifiers to avoid overfitting and reduce computational consumption.

[0090] The weights of the C-RNN model are initialized using the He normal distribution for convolutional layer weights to promote stable gradient propagation in the early stages of training. All layers with bias terms (such as convolutional layers and fully connected layers) are initialized to zero, and layers without bias terms (such as pooling layers) are not initialized.

[0091] In terms of training strategy, a 50-round training limit was set to balance effectiveness and computational efficiency. The Adam algorithm was used for optimization, with an initial learning rate of 0.001. Dynamic adjustment strategies were implemented during training to adapt to the training needs at different stages.

[0092] The data was divided using a 5-fold cross-validation method. In each iteration, 20% of the pixels in the dataset were randomly selected as the training set, 70% as the test set, and the remaining 10% as the validation set.

[0093] Step 4: Use three commonly used indicators in this field to comprehensively evaluate the results, namely overall accuracy (OA), average accuracy (AA) and Kappa coefficient.

[0094] Three typical dual-branch spatial-spectral fusion models (2D-CNN-BiRNN-AdaBoost, 2D-1DCNN-AdaBoost, and 2D-LSTM-AdaBoost) were selected to compare with the C-RNN-AdaBoost model of the present invention to verify the effectiveness and innovation of the GRU gating mechanism. The overall accuracy (OA), average accuracy (AA), and Kappa coefficient of the proposed model are higher (as shown in Table 1). Figure 6 The results (shown in Figure 2) show that the proposed model has a higher smoothness in rock classification and fewer misclassified pixels, which verifies its efficiency and robustness in complex rock classification tasks.

[0095] Table 1 Comparison of model classification accuracy under different RNN branches

[0096] Model OA (%) AA (%) Kappa×100 Parameter quantity 2D-BRNN-AdaBoost 83.914±1.579 83.350±1.700 83.70±1.6 16342941 2D-LSTM-AdaBoost 89.796±2.842 89.298±3.194 89.70±2.9 8229085 2D-1DCNN-AdaBoost 91.118±5.227 90.670±5.461 91.10±5.3 36670263 C-RNN-AdaBoost 92.554±7.442 92.250±7.861 92.50±7.5 1012632

[0097] Three mainstream hyperspectral image classification algorithms (3D-CNN, Bi-CLSTM, SSRN, DenseNet) were selected for comparative analysis with the C-RNN-AdaBoost of the present invention to verify the effectiveness of the rock classification algorithm. According to the data in Table 2, the model proposed in this paper showed the highest classification accuracy in the 81-class rock classification task. The classification effect diagram of the four comparison models and the proposed model is shown in Figure 2. Figure 7 shown.

[0098] Table 2 Classification results of different models on 81 types of rock datasets

[0099] Model OA (%) AA (%) Kappa×100 3DCNN 87.129±6.439 88.235±6.873 87.10±6.5 Bi-CLSTM 89.476±3.316 89.038±3.128 88.20±3.0 SSRN 91.012±4.734 91.989±4.325 90.90±4.9 DenseNet 90.301±3.213 90.991±3.052 90.80±3.1 C-RNN-AdaBoost 92.554±7.442 92.250±7.861 92.50±7.5

[0100] In summary, the present invention combines the spatial feature extraction module, the spectral feature extraction module and the AdaBoost optimization module to build a hyperspectral rock classification model based on deep ensemble learning. The results show that this model achieved 92.55%, 92.25% and 92.50% in overall accuracy (OA), average accuracy (AA) and Kappa coefficient, respectively, which is significantly better than the mainstream model, demonstrating the effectiveness of multimodal feature fusion and ensemble learning strategies, and effectively solving the shortcomings of traditional methods in terms of efficiency, data dimensionality disaster and small sample overfitting. At the same time, it supports multimodal data fusion and mobile terminal deployment, providing intelligent and high-precision solutions for geological mapping, mineral exploration and other fields, with both theoretical innovation and engineering application value.

[0101] Of course, the above examples of the present invention are only intended to illustrate the implementation process of the present invention and are not intended to limit the embodiments of the present invention. Those skilled in the art may make corresponding changes and modifications based on the present invention, but such changes and modifications should fall within the scope of protection of the claims appended hereto.

Claims

1. A hyperspectral rock classification method based on deep ensemble learning algorithm, characterized by include: Firstly, a hyperspectral dataset of 81 types of rock samples was obtained from the hyperspectral rock standard database. After dimensionality reduction using principal component analysis, it was divided into three-dimensional cubes and input into a 2D convolutional neural network as spatial features. At the same time, the central pixel block after dimensionality reduction was input into a gated recurrent unit as spectral features. Then, the 2D convolutional neural network and the gated recurrent unit were connected in series, and a fully connected layer was introduced to fuse spatial features and spectral features. The model performance was optimized with the AdaBoost algorithm. Subsequently, the hyperspectral dataset after principal component analysis dimensionality reduction was divided into training and test sets according to a certain ratio, and the hyperspectral rock classification model was trained in batches. During the training process, the learning rate and weak classifier weights were dynamically adjusted until the model converged. Finally, the classification effect of the trained model was evaluated using the test set, and a classification result graph was generated to comprehensively test and demonstrate the classification ability of the model.

2. The hyperspectral rock classification method based on deep ensemble learning algorithm according to claim 1 is characterized in that: The hyperspectral rock classification model includes a spatial feature extraction module, a spectral sequence feature extraction module, and an AdaBoost optimization module; Spatial Feature Extraction Module: This module streamlines the number of convolution channels and layers, constructing multiple customized 2D convolutional layers with different parameters and kernel sizes. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function, while omitting the pooling layer to maintain high spatial resolution. These 2D convolutional layers efficiently capture multi-scale spatial structural features from the raw hyperspectral image data after PCA dimensionality reduction, generating Spatial Feature Map I. Then these spatial feature maps I are fused to obtain spatial feature map II that integrates multi-scale spatial features, in preparation for subsequent fusion with spectral features; When fused with the spectral features, the spatial feature map II is input into the fully connected layer; Spectral sequence feature extraction module: This module combines 1D-CNN and GRU. 1D-CNN processes spectral data using a specific convolution kernel and adjusting the step size and padding parameters. After batch normalization and ReLU activation, it generates a preliminary spectral feature map I. The GRU layer uses a reset gate and update gate mechanism to stack multiple layers to capture spectral features at different time scales, generating a spectral feature map II containing rich spectral response patterns. The spectral feature map II is then integrated and optimized using batch normalization and the ReLU activation function to obtain the spectral feature map III. Finally, the spectral feature map III is input into the fully connected layer and deeply fused with the spatial feature map II of the 2D-CNN branch. The neurons in the fully connected layer are connected to the elements of the two feature maps to obtain a fused feature map, which improves the classification performance of hyperspectral rock images. The AdaBoost optimization module plays a role in dynamically adjusting sample weights and model decision weights in the hyperspectral rock classification model. In each iteration, AdaBoost adjusts sample weights based on the current classification results of the fully connected fusion feature map, increasing the weights of misclassified samples. At the same time, AdaBoost enables the model to better cope with noise and outliers when processing the fusion feature map. For samples with "same spectrum, different objects" or "same object, different spectrum", it adjusts the focus on relevant features and improves model robustness. In addition, the AdaBoost optimization module calculates and adjusts the weights of weak classifiers. In the process of weighted linear superposition and integration of weak classifiers, it highlights the role of important spectral features in the fusion feature map, suppresses interference, and improves classification accuracy and reliability.

3. The hyperspectral rock classification method based on deep ensemble learning according to claim 2 is characterized in that: As a variant of LSTM, GRU has a simpler gating mechanism, which only contains update gate and reset gate. and the hidden states of the previous time steps In case of a fault, reset gate R t and update gate Z t The calculation formula is as follows: R t =sigmoid(X t W xr +H t-1 W hr +b r ) (1) Z t =sigmoid(X t W xz +H t-1 W hz +b z ) (2) Where n is the number of samples, d is the number of inputs, and h is the hidden unit dimension; In formula (1), W xr represents the weight matrix from the input layer to the reset gate; W hr represents the recursive weight from the hidden state to the reset gate; b r is the bias vector for resetting the gate; In formula (2), W xz Represents the projection weight input to the update gate; W hz represents the recursive weight from the hidden state to the update gate; b z To update the gate bias; The Sigmoid function is used as an activation function to map values between 0 and 1, controlling the gate mechanism in the GRU; this makes the reset gate R t and update gate Z t The output value is in the range [0,1]; GRU calculates the candidate hidden state To assist in updating the hidden state, this operation is performed by performing element-wise multiplication between the output of the reset gate at the current time step and the hidden state of the previous time step; specifically, the candidate hidden state at time step t The calculation formula is as follows: In formula (3), W xh Represents the projection weight of the input to the candidate state; W hh represents the recursive weight of the hidden state after reset; b h is the candidate state bias, and ⊙ represents element-wise multiplication.

4. The hyperspectral rock classification method based on deep ensemble learning according to claim 2 is characterized in that The AdaBoost optimization module implements a dual enhancement mechanism by dynamically adjusting sample weights and model decision weights. It also iteratively adjusts the sample weight distribution to force the model to continuously focus on difficult samples. The specific execution process is as follows: First, initialize the sample weights; assuming that the dataset D contains M samples, the initial weight of each sample is equal to That is, the weight of the i-th sample in the first round of iteration; Then use the current sample weight to train a weak classifier c t (x), and calculate the weighted error rate r of the weak classifier on the training set t , formula (4) is as follows: In formula (4), I(·) is an indicator function, which takes the value of 1 when the prediction is wrong and 0 otherwise; i represents the eigenvector of the i-th sample, so c t (x i ) is the weak classifier for the tth round of sample x i The predicted value, y i That is, the input data represents the label of the i-th sample, that is, the output target that the model needs to predict; According to the weighted error rate r t Calculate the weight a of the weak classifier t , as shown in formula (5): Weight a t Reflects the importance of the weak classifier in the final decision, the weighted error rate r t The lower it is, the higher the weight; Then, update the sample weights so that the weights of misclassified samples increase, as shown in formula (6): In each iteration of the AdaBoost algorithm, the sample weights need to be normalized after being updated, that is, each sample weight value is divided by the current total weight so that the sum of the normalized weights is constant to 1; When all the preset iteration rounds are completed, the algorithm will calculate the corresponding weight a of each weak classifier according to its error rate during training. t , and finally all weak classifiers are integrated into a strong classifier by weighted linear superposition; its mathematical expression is shown in formula (7): Where sign(·) is a sign function used to determine the final classification result; T represents the total number of iterations, that is, the total number of weak classifiers used in the training process.

5. The hyperspectral rock classification method based on deep ensemble learning according to claim 1 or 2, characterized in that The model is trained in batches until convergence is stable, which includes: The AdaBoost algorithm configuration uses 10 decision tree stumps as base classifiers and it is determined that the best balance is achieved when using 3 classifiers; In terms of training strategy, a training limit of 50 rounds is set; the Adam algorithm is used in the optimization process, the initial learning rate is set to 0.001, and a dynamic adjustment strategy is implemented during the training process; The data was divided using a 5-fold cross-validation method; in each iteration, 20% of the pixels in the dataset were randomly selected as the training set, 70% as the test set, and the remaining 10% as the validation set.

Images