An underground winter bamboo shoot recognition method based on deep learning multi-domain feature fusion
By employing a deep learning-based multi-domain feature fusion method, combined with short-time Fourier transform, autocorrelation analysis, and wavelet transform, a multi-domain feature set is constructed. Utilizing the RES-ICSE-CBAM convolutional neural network model, the problem of low efficiency in traditional manual bamboo shoot exploration is solved, enabling accurate identification and efficient development of winter bamboo shoots.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FUJIAN AGRI & FORESTRY UNIV
- Filing Date
- 2026-04-24
- Publication Date
- 2026-07-21
AI Technical Summary
Traditional manual bamboo shoot detection is inefficient and destructive. Existing detection technologies have limited feature extraction and insufficient identification accuracy, making it difficult to accurately identify winter bamboo shoots.
A deep learning-based multi-domain feature fusion method is adopted. Signal features are extracted through short-time Fourier transform, autocorrelation analysis and wavelet transform to construct a multi-domain feature set. The RES-ICSE-CBAM convolutional neural network model is used for feature learning and fusion. Finally, the classification results are output through the stacking strategy.
This improved the accuracy and efficiency of winter bamboo shoot identification, reduced damage to bamboo forests, and enabled the accurate identification and efficient development of winter bamboo shoot resources.
Smart Images
Figure CN122432991A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of underground target recognition and signal processing technology, specifically relating to an underground winter bamboo shoot recognition method based on deep learning multi-domain feature fusion. Background Technology
[0002] Winter bamboo shoots are products of the differentiation and development of lateral buds on the underground stems of moso bamboo. As an important economic resource in my country's bamboo forest ecosystem, they are rich in dietary fiber, various vitamins, protein, total sugar, and other functional nutrients, as well as minerals such as potassium, calcium, and phosphorus, possessing both high nutritional value and health benefits. Furthermore, winter bamboo shoots are low in calories and virtually fat-free, making them an excellent choice for a low-fat diet and a popular green and healthy food among consumers. Due to their outstanding nutritional value, the market demand for winter bamboo shoots continues to rise.
[0003] Winter bamboo shoots are mostly distributed in the soil layer 3-20cm underground. Due to the concealed nature of their growing environment, their harvesting has long relied on the traditional manual soil-turning and exploration method: observing the growth of bamboo branches and leaves, and searching around the base of the bamboo along the rhizomes. This traditional method is not only labor-intensive, inefficient, and prone to missed detection, but also severely damages the bamboo forest's root system and soil structure through indiscriminate digging, threatening the stability of the bamboo forest ecosystem. Furthermore, it suffers from drawbacks such as strong reliance on manual labor and a lack of standardization and large-scale operation capabilities. At the same time, unharvested winter bamboo shoots will rot in the soil, interfering with the absorption and cycling of nutrients in the rhizomes, thus affecting the growth and development of bamboo the following year and hindering the sustainable management of bamboo forests. Therefore, achieving accurate identification of winter bamboo shoots is crucial to improving the efficiency of winter bamboo shoot resource development and ensuring the sustainable development of bamboo forests.
[0004] In recent years, the application of ground-penetrating radar (GPR) technology in the field of underground target detection has developed rapidly, and related technologies have accumulated relatively rich research results in the detection of underground root systems and other targets. However, directly transferring this technology to the scenario of winter bamboo shoot detection still faces many difficulties. In the actual bamboo forest ecological environment, winter bamboo shoots have different growth forms, and the soil is mixed with various interfering substances such as tree roots and bamboo rhizomes. Their dielectric constants are similar to those of winter bamboo shoots, which can easily confuse the echo signals, causing the characteristic boundaries between the target signal and the interference signal to be blurred, making it difficult to effectively distinguish them. If relying solely on the traditional method of manually extracting single-dimensional signal features, it is difficult to achieve ideal results due to insufficient feature representation capabilities and the limitations of human subjective judgment.
[0005] To address the challenge of classifying and identifying complex underground target signals, machine learning and deep learning technologies have been widely applied in signal processing in recent years, becoming important tools for solving complex signal feature extraction and recognition problems, and demonstrating excellent performance in multi-class signal classification scenarios. However, existing signal classification and recognition methods are mostly limited to feature transformation in a single domain, especially time-frequency domain feature analysis. These traditional strategies fail to fully leverage the synergistic and complementary advantages of multi-domain signal features and neglect the crucial role of multi-domain feature fusion in improving target classification accuracy. In conclusion, developing a target recognition and classification method that can integrate multi-domain signal information to achieve accurate identification of winter bamboo shoots has become an urgent need for the upgrading and technological innovation of the bamboo industry. Summary of the Invention
[0006] The purpose of this invention is to solve the problems of low efficiency and high damage of traditional manual bamboo shoot detection, as well as the single feature extraction and insufficient recognition accuracy of existing detection technologies. It provides a method for identifying underground winter bamboo shoots based on deep learning multi-domain feature fusion.
[0007] To achieve the above objectives, the technical solution of the present invention is: a method for identifying underground winter bamboo shoots based on deep learning multi-domain feature fusion, comprising:
[0008] Collect echo signals from underground targets, including winter bamboo shoot targets, bamboo rhizomes, and soil background signals;
[0009] The echo signals are preprocessed to construct an underground target signal recognition dataset;
[0010] Based on the underground target signal recognition dataset, the time-frequency features, autocorrelation domain features, and wavelet domain features of the signal are extracted by short-time Fourier transform, autocorrelation analysis, and wavelet transform, respectively, and a multi-domain feature set is constructed.
[0011] Three sets of RES-ICSE-CBAM convolutional neural network models with the same structure but different input features were constructed to learn time-frequency features, autocorrelation domain features and wavelet domain features, respectively.
[0012] The Stacking multi-domain feature fusion strategy is adopted, which uses the outputs of three sets of RES-ICSE-CBAM convolutional neural network models as meta-features, inputs them into the neural network meta-model for fusion, and outputs the final classification result.
[0013] Furthermore, a high-frequency electromagnetic wave winter bamboo shoot detector is used to collect echo signals; the detector includes a control module, a transmission link and a receiving link. The main control unit controls the development board to generate high-frequency signals, which are radiated to the target area via the transmitting antenna of the transmission link, and the echo signals are received by the receiving antenna of the receiving link.
[0014] Furthermore, the short-time Fourier transform divides the signal into multiple segments by introducing a sliding window and performs a Fourier transform on each segment. The number of Fourier transform points is set to 1024, the sliding window length is 512, and the output result is converted into a two-dimensional time-frequency image after normalization.
[0015] Furthermore, the autocorrelation analysis extracts features by calculating the similarity between the signal and its delayed version, and the output is converted into a two-dimensional autocorrelation domain image after normalization.
[0016] Furthermore, the wavelet transform employs the "db4" wavelet basis function and decomposes the signal through three layers of wavelets to capture transient changes and non-stationary characteristics; the output result is then normalized and converted into a two-dimensional wavelet domain image.
[0017] Furthermore, the RES-ICSE-CBAM convolutional neural network model includes a RES module, an ICSE module, and a CBAM module connected in sequence;
[0018] The RES module includes a main path and a skip connection path. The features of the two paths are added and fused through residual connections, and finally output through the ReLU activation function.
[0019] The ICSE module includes an IC submodule and an SE submodule. The IC submodule adopts a three-branch parallel architecture, extracting features from different receptive fields through 3×3 convolution, 1×1 convolution + 3×3 convolution, and 1×1 convolution + 5×5 convolution, respectively, and then concatenating the features and fusing them through 1×1 convolution to reduce dimensionality. The SE submodule performs global average pooling on the aggregated features and learns channel weights through two levels of fully connected layers.
[0020] The CBAM module includes a channel attention submodule and a spatial attention submodule.
[0021] Furthermore, the processing procedure of the CBAM module is expressed as follows:
[0022]
[0023]
[0024] In the formula, The original input features, The channel weights are generated after the original input features are processed through channel attention. The spatial weights are generated after the original input features are processed by spatial attention, and ⊗ indicates element-wise multiplication.
[0025] Furthermore, the channel attention submodule employs a dual pooling strategy of global average pooling and global max pooling. After inputting the pooling results into a multilayer perceptron with shared parameters, the channel weights are generated through a sigmoid activation function. The spatial attention submodule performs average pooling and max pooling on the original input features along the spatial dimension, concatenates the results, and generates spatial weights through a 3×3 convolution and a sigmoid activation function. The original input features are simultaneously input into the channel attention submodule and the spatial attention submodule, and after generating corresponding weights, they are multiplied sequentially by the channel weights and spatial weights to output the optimized feature map.
[0026] Furthermore, each RES-ICSE-CBAM convolutional neural network model initially extracts preliminary features through a convolutional layer containing 64 3×3 convolutional kernels. Subsequently, the feature map is input into a three-level module group composed of RES-ICSE-CBAM blocks. Max pooling layers are introduced after the first two levels of module groups, and global average pooling is used to compress the features into vectors after the last level of module group. After processing by fully connected layers and Dropout layers, the classification probability is output by the softmax function.
[0027] Furthermore, the neural network meta-model consists of fully connected layers, Dropout layers, and ReLU activation functions; the Stacking strategy concatenates the classification probability vectors output by the base models of the three RES-ICSE-CBAM convolutional neural network models for each sample, and uses them as new features to input into the neural network meta-model for training and prediction.
[0028] Compared with the prior art, the present invention has the following beneficial effects:
[0029] This invention introduces the concept of "multi-domain feature fusion" into the specific agricultural application scenario of winter bamboo shoot detection. Unlike previous studies that may have focused on single signal processing methods, this research collaboratively analyzes the detection signal from multiple information dimensions, including the time-frequency domain, autocorrelation domain, and wavelet domain, providing a more comprehensive new perspective on feature representation and helping to reveal the essential attributes of the winter bamboo shoot target more profoundly. This invention proposes a hierarchical intelligent recognition framework of "benchmark model + ensemble learning". In the design of the benchmark CNN model, modules such as residual structure and attention mechanism are integrated. The residual structure alleviates the gradient vanishing problem in deep network training, while the attention mechanism can adaptively calibrate the feature response, making the model pay more attention to the key information regions related to winter bamboo shoot recognition and suppressing background noise interference, thereby improving the precision of feature extraction and the model's discriminative ability; furthermore, through the ensemble learning strategy, intelligent fusion and complementarity of multi-model prediction results are achieved at the decision-making level. Attached Figure Description
[0030] Figure 1 This is a schematic diagram of the bamboo shoot detector's detection principle in a specific embodiment of the present invention;
[0031] Figure 2 This is a two-dimensional time-frequency domain feature map in a specific embodiment of the present invention;
[0032] Figure 3 This is a two-dimensional autocorrelation domain feature map in a specific embodiment of the present invention;
[0033] Figure 4 This is a two-dimensional wavelet domain feature map in a specific embodiment of the present invention;
[0034] Figure 5 This is a diagram of the RES module of the baseline model in a specific embodiment of the present invention;
[0035] Figure 6 This is an ICSE module diagram of the baseline model in a specific embodiment of the present invention;
[0036] Figure 7 This is a CBAM module diagram of the baseline model in a specific embodiment of the present invention;
[0037] Figure 8 This is a multi-domain feature fusion strategy diagram in a specific embodiment of the present invention. Detailed Implementation
[0038] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.
[0039] This invention provides a method for identifying underground winter bamboo shoots based on deep learning multi-domain feature fusion, comprising:
[0040] Collect echo signals from underground targets, including winter bamboo shoot targets, bamboo rhizomes, and soil background signals;
[0041] The echo signals are preprocessed to construct an underground target signal recognition dataset;
[0042] Based on the underground target signal recognition dataset, the time-frequency features, autocorrelation domain features, and wavelet domain features of the signal are extracted by short-time Fourier transform, autocorrelation analysis, and wavelet transform, respectively, and a multi-domain feature set is constructed.
[0043] Three sets of RES-ICSE-CBAM convolutional neural network models with the same structure but different input features were constructed to learn time-frequency features, autocorrelation domain features and wavelet domain features, respectively.
[0044] The Stacking multi-domain feature fusion strategy is adopted, which uses the outputs of three sets of RES-ICSE-CBAM convolutional neural network models as meta-features, inputs them into the neural network meta-model for fusion, and outputs the final classification result.
[0045] The following is a detailed implementation process of the present invention.
[0046] A method for identifying underground winter bamboo shoots based on deep learning multi-domain feature fusion includes the following steps:
[0047] 1) A self-developed high-frequency electromagnetic wave winter bamboo shoot detector is used to collect echo signals corresponding to three types of targets: scenes containing fresh winter bamboo shoots, scenes containing bamboo rhizomes but no winter bamboo shoots, and pure soil scenes containing weeds but no plant roots. Correspondingly, winter bamboo shoot target signals, bamboo rhizome target signals, and soil background signals are obtained. The detector principle is as follows: Figure 1 As shown, the detector consists of a control module, a transmit link, and a receive link. The Jetson Nano main control unit controls the B210 development board to generate high-frequency signals, which are then directionally radiated to the target area by the transmitting antenna. When the high-frequency signal comes into contact with the medium, it generates echo signals. These echo signals are received by the receiving antenna and transmitted to the receiving device for storage and subsequent analysis.
[0048] 2) Preprocess the signals obtained in step 1) to construct an underground target signal recognition dataset. The preprocessing process includes signal length standardization, zero-padding and truncation. At the same time, Gaussian noise (mean of 0, variance of 0.01) is added for data augmentation to form an underground target recognition dataset, ensuring that the number of samples in the three classes is balanced.
[0049] 3) Based on the dataset constructed in step 2), time-frequency features, autocorrelation domain features, and wavelet domain features of the signal are extracted through short-time Fourier transform, autocorrelation analysis, and wavelet transform, respectively, thereby constructing a multi-domain feature set. Specifically, the short-time Fourier transform divides the signal into multiple segments by introducing a sliding window, and performs a Fourier transform on each segment, thus simultaneously preserving the time and frequency information of the signal. The number of Fourier transform points is set to 1024, and the sliding window length is 512. The output result is normalized and converted into a two-dimensional time-frequency image, as shown below. Figure 2 As shown. By Figure 2As can be seen, all three types of signals exhibit two bright, concentrated energy stripes in the 0.5 GHz and 1.5 GHz frequency bands, indicating that these two frequency bands are the main energy distribution ranges after the interaction between electromagnetic waves and underground targets. In other frequency bands, the characteristics of the three types of signals differ significantly: in the scenario containing bamboo shoots, the brightness of the signal stripes shows a complex gradient change, reflecting its drastic energy fluctuations; in the scenario containing bamboo rhizomes, the degree of change in stripe brightness is less, and the energy fluctuation amplitude is weaker than in the bamboo shoot scenario; in the pure soil scenario, the stripe brightness remains basically stable, and the overall energy distribution is gentle with minimal fluctuations. The underlying reason is that the water content of winter bamboo shoots and bamboo rhizomes is higher than that of soil, and their dielectric constant is also greater than that of soil. This makes the scattering and reflection of electromagnetic waves within them more complex, thus causing significant energy fluctuations. The irregular shape of winter bamboo shoots and the uneven distribution of internal moisture further exacerbate the randomness of electromagnetic wave scattering / reflection, resulting in a rich variety of energy change characteristics. Bamboo rhizomes have a slender structure and a relatively uniform moisture distribution. The intensity and form of their effect on electromagnetic waves are different from those of winter bamboo shoots, making the degree of energy fluctuation between that of winter bamboo shoots and soil. In contrast, the dielectric constant of pure soil is stable, the medium composition is simple, and the electromagnetic wave scattering and reflection are simple. Therefore, the energy distribution is smooth, the fluctuation amplitude is very small, and the time-frequency characteristics are stable.
[0050] Autocorrelation analysis, by calculating the similarity between a signal and its delayed version, can effectively reflect the periodicity and repetitiveness of a signal. The output is then normalized and converted into a two-dimensional autocorrelation domain image, such as... Figure 3 As shown. By Figure 3 It can be seen that the signal stripes are darkest in the scene containing winter bamboo shoots, indicating that the signal energy loss is most severe in this scene; the signal stripes are relatively bright in the scene containing bamboo rhizomes, corresponding to a slightly lower energy loss than in the winter bamboo shoot scene; the signal stripes are brightest in the pure soil scene, with the lowest degree of energy loss.
[0051] Wavelet transform can effectively capture transient changes and non-stationary characteristics in signals through multi-scale analysis. Using the "db4" wavelet basis function, a three-level wavelet decomposition is performed, and the output is normalized and converted into a two-dimensional wavelet domain image, such as... Figure 4 As shown. By Figure 4 visible, Figure 4 In the left-middle image, where winter bamboo shoots are present, the signal energy is concentrated in the high-scale range of scale level 23-26. Figure 4 In the intermediate image containing the bamboo whip scene, the signal energy is also concentrated in the high-scale range of 23-26, but compared with... Figure 4 Compared to the left image, the bright area in the bamboo whip scene is significantly larger, and the color intensity is higher across the entire scale, indicating that the overall signal energy in the bamboo whip scene is higher than that in the winter bamboo shoot scene. This is because although the energy distribution patterns in the wavelet domain are similar, the dielectric constant of winter bamboo shoots is slightly greater than that of bamboo whips, which accelerates the attenuation of electromagnetic wave energy, thereby reducing the overall signal energy intensity. In contrast, Figure 4The color distribution of the signal in the pure soil scenario (middle right image) differs significantly from the previous two scenarios: the energy is mainly concentrated in the scale range of 14-21, and fluctuates dramatically over time, which is the result of random scattering of electromagnetic waves caused by the non-uniformity of the soil itself; in other scale ranges, the signal energy in the pure soil scenario remains relatively stable. These differences can serve as a key basis for accurately distinguishing winter bamboo shoots, bamboo rhizomes, and soil.
[0052] 4) Construct three sets of RES-ICSE-CBAM parallel convolutional neural network models. The three models have the same network structure and differ only in their input features. Each model is optimized for a specific type of feature. The specific structure of each module is as follows:
[0053] The RES module introduces skip connections, enabling neural networks to learn the residual mapping between inputs and outputs, effectively addressing the vanishing gradient and network degradation problems. For example... Figure 5 As shown, the input features first pass through the main path, sequentially through a 3×3 convolution, a batch normalization module, and a ReLU activation function, followed by another 3×3 convolution and batch normalization. In the skip connection path, if the number of input and output feature channels does not match, a 1×1 convolution is used to adjust the feature dimensions; if the number of channels is the same, no adjustment is made. The two paths are joined and fused through residual connections, and finally output through a ReLU activation function.
[0054] The ICSE module employs a parallel design of multi-size convolutional kernels and pooling layers, enabling multi-scale feature extraction within a single network layer, effectively overcoming the limitation of the single receptive field in traditional sequential convolutional structures. Furthermore, it compresses spatial information through global average pooling and utilizes fully connected layers to learn inter-channel dependencies, achieving adaptive calibration of feature channels and significantly improving the model's feature selection capability with minimal computational cost. Figure 6 As shown, the IC submodule adopts a three-branch parallel architecture, extracting features from different receptive fields through "3×3 convolution", "1×1 convolution + 3×3 convolution", and "1×1 convolution + 5×5 convolution", respectively. The concatenated features are then fused and dimensionality reduced through 1×1 convolution. The SE submodule further performs global average pooling on the aggregated features to achieve dimensionality reduction, and dynamically learns channel weights through two levels of fully connected layers to enhance the expressive power of key features.
[0055] CBAM module: This is a lightweight attention mechanism module that enhances feature representation capabilities through dual channel and spatial attention mechanisms. This invention improves upon the original CBAM by employing a parameter-sharing MLP to reduce the computational cost of the channel attention module and simplifies the spatial attention module using a single-layer 3×3 convolution. The overall attention process can be summarized as follows:
[0056]
[0057]
[0058] In the formula, The original input features, The channel weights are generated after the original input features are processed through channel attention. Spatial weights are generated after spatial attention is applied to the original input features.
[0059] The improved CBAM module first applies channel attention, then spatial attention. This represents a significant innovation compared to traditional structures: the spatial attention module directly uses the original feature map as input, skipping the output stage of channel attention. This optimized design effectively avoids feature redundancy and information loss. Figure 7 As shown, the channel attention module employs a dual-pooling strategy, using global average pooling and global max pooling to extract statistical information from the original features. Feature transformation is then performed via a parameter-sharing multilayer perceptron (MLP), and finally, channel weights are generated using a sigmoid activation function. The spatial attention module performs average pooling and max pooling operations along the spatial dimension of the original input features. The results of the two pooling operations are concatenated and then passed through a 3×3 convolutional layer and a sigmoid activation function to generate spatial weights. The original input features are simultaneously fed into the dual-channel attention module, where corresponding weights are generated and then multiplied sequentially by the channel weights and spatial weights to output the optimized feature map.
[0060] 5) The Stacking multi-domain feature fusion strategy generates meta-features through a base model, and then uses the meta-model to fuse the prediction information of multiple base models, giving full play to the advantages of each model and improving the overall recognition performance. For example... Figure 8 As shown, features of three target signals—winter bamboo shoots, bamboo rhizomes, and soil—are extracted using short-time Fourier transform, autocorrelation analysis, and wavelet transform, respectively, and converted into two-dimensional feature images. These images are then input into three basic models equipped with a hybrid attention mechanism. Initially, the models perform preliminary feature extraction using convolutional layers with 64 3×3 kernels. The feature maps are then input into a three-level module group consisting of RES-ICSE-CBAM blocks for iterative optimization. The first two levels of the module group introduce max pooling layers for downsampling during feature processing. The last level of the module group compresses the features into vectors using global average pooling. After processing by fully connected layers and Dropout regularization, the softmax activation function outputs the classification result. After the basic models corresponding to the three types of features are trained, a stacking fusion strategy is adopted. The output probabilities of each basic model are used as new features and input into a neural network meta-model composed of fully connected layers, Dropout layers, and ReLU activation functions. Early stopping and learning rate decay strategies are combined to complete training optimization, ultimately outputting the fused classification prediction result.
[0061] 6) To further verify the effectiveness of the proposed fusion strategy, STFT, Autocorr, and Wavelet features were combined in pairs and fused with the three domains respectively. These were used as model inputs to conduct comparative experiments. The specific evaluation indicators are shown in Table 1.
[0062] Table 1 Evaluation metrics for input models with different features
[0063]
[0064] As shown in Table 1, the three-domain fusion feature (STFT+Autocorr+Wavelet) exhibits the best performance across all evaluation metrics, with accuracy, precision, F1 score, and AUC reaching 0.8441, 0.8487, 0.8443, and 0.9243, respectively, outperforming schemes that combine any two features pairwise. This result fully demonstrates that fusing multi-domain features can more comprehensively characterize signal properties, effectively compensating for the limitations of single-domain feature representation capabilities, and thus significantly improving the model's classification performance for underground targets.
[0065] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.
Claims
1. A method for identifying underground winter bamboo shoots based on deep learning multi-domain feature fusion, characterized in that, include: Collect echo signals from underground targets, including winter bamboo shoot targets, bamboo rhizomes, and soil background signals; The echo signals are preprocessed to construct an underground target signal recognition dataset; Based on the underground target signal recognition dataset, the time-frequency features, autocorrelation domain features, and wavelet domain features of the signal are extracted by short-time Fourier transform, autocorrelation analysis, and wavelet transform, respectively, and a multi-domain feature set is constructed. Three sets of RES-ICSE-CBAM convolutional neural network models with the same structure but different input features were constructed to learn time-frequency features, autocorrelation domain features and wavelet domain features, respectively. The Stacking multi-domain feature fusion strategy is adopted, which uses the outputs of three sets of RES-ICSE-CBAM convolutional neural network models as meta-features, inputs them into the neural network meta-model for fusion, and outputs the final classification result.
2. The method for identifying underground winter bamboo shoots based on deep learning multi-domain feature fusion according to claim 1, characterized in that, The echo signal is collected using a high-frequency electromagnetic wave winter bamboo shoot detector. The detector includes a control module, a transmission link and a receiving link. The main control unit controls the development board to generate a high-frequency signal, which is radiated to the target area through the transmitting antenna of the transmission link, and the echo signal is received by the receiving antenna of the receiving link.
3. The method for identifying underground winter bamboo shoots based on deep learning multi-domain feature fusion according to claim 1, characterized in that, The short-time Fourier transform divides the signal into multiple segments by introducing a sliding window and performs a Fourier transform on each segment. The number of Fourier transform points is set to 1024, the sliding window length is 512, and the output result is converted into a two-dimensional time-frequency image after normalization.
4. The method for identifying underground winter bamboo shoots based on deep learning multi-domain feature fusion according to claim 1, characterized in that, The autocorrelation analysis extracts features by calculating the similarity between the signal and its delayed version, and the output is converted into a two-dimensional autocorrelation domain image after normalization.
5. The method for identifying underground winter bamboo shoots based on deep learning multi-domain feature fusion according to claim 1, characterized in that, The wavelet transform uses the "db4" wavelet basis function and is decomposed through three layers of wavelets to capture the transient changes and non-stationary characteristics of the signal; the output result is normalized and then converted into a two-dimensional wavelet domain image.
6. The method for identifying underground winter bamboo shoots based on deep learning multi-domain feature fusion according to claim 1, characterized in that, The RES-ICSE-CBAM convolutional neural network model includes a RES module, an ICSE module, and a CBAM module connected in sequence. The RES module includes a main path and a skip connection path. The features of the two paths are added and fused through residual connections, and finally output through the ReLU activation function. The ICSE module includes an IC submodule and an SE submodule. The IC submodule adopts a three-branch parallel architecture, extracting features from different receptive fields through 3×3 convolution, 1×1 convolution + 3×3 convolution, and 1×1 convolution + 5×5 convolution, respectively, and then concatenating the features and fusing them through 1×1 convolution to reduce dimensionality. The SE submodule performs global average pooling on the aggregated features and learns channel weights through two levels of fully connected layers. The CBAM module includes a channel attention submodule and a spatial attention submodule.
7. The method for identifying underground winter bamboo shoots based on deep learning multi-domain feature fusion according to claim 6, characterized in that, The processing procedure of the CBAM module is represented as follows: In the formula, The original input features, The channel weights are generated after the original input features are processed through channel attention. The spatial weights are generated after the original input features are processed by spatial attention, and ⊗ indicates element-wise multiplication.
8. The method for identifying underground winter bamboo shoots based on deep learning multi-domain feature fusion according to claim 7, characterized in that, The channel attention submodule adopts a dual pooling strategy of global average pooling and global max pooling. The pooling result is input into a multilayer perceptron with shared parameters, and then channel weights are generated by the sigmoid activation function. The spatial attention submodule performs average pooling and max pooling on the original input features along the spatial dimension, and concatenates the results before generating spatial weights through a 3×3 convolution and a sigmoid activation function. The original input features are simultaneously passed to the channel attention submodule and the spatial attention submodule. After generating corresponding weights, they are multiplied by the channel weights and spatial weights in sequence, and the optimized feature map is output.
9. The method for identifying underground winter bamboo shoots based on deep learning multi-domain feature fusion according to claim 1, characterized in that, Each RES-ICSE-CBAM convolutional neural network model initially extracts preliminary features through a convolutional layer with 64 3×3 convolutional kernels. The feature map is then input into a three-level module group consisting of RES-ICSE-CBAM blocks. Max pooling layers are introduced after the first two levels of module groups, and global average pooling is used to compress the features into vectors after the last level of module groups. After processing by fully connected layers and Dropout layers, the classification probability is output by the softmax function.
10. The method for identifying underground winter bamboo shoots based on deep learning multi-domain feature fusion according to claim 1, characterized in that, The neural network metamodel consists of a fully connected layer, a Dropout layer, and a ReLU activation function. The Stacking strategy concatenates the classification probability vectors output by the base models of the three RES-ICSE-CBAM convolutional neural network models for each sample, and uses this as a new feature input to the neural network metamodel for training and prediction.