A soybean phenological stage automatic identification method based on unmanned aerial vehicle multi-modal data

Abstract

The present application relates to the field of agricultural remote sensing and computer vision, and particularly relates to a soybean phenological stage automatic identification method based on unmanned aerial vehicle multi-modal data, which comprises the following technical scheme: multi-modal information obtained by an unmanned aerial vehicle, including image data and tabular feature data, is used to construct an innovative dynamic gating fusion deep learning framework, a dynamic weight adjustment mechanism is introduced to realize adaptive fusion of multi-source features and time series feature modeling, so that high-precision and automatic classification of soybean key phenological stages is realized; the method not only has excellent classification precision and an innovative intelligent fusion mechanism, but also has significant collaborative error correction capability, and has high throughput and real-time monitoring capability, and has complete technical verification and scalability.

Description

Technical Field

[0001] This invention relates to the fields of agricultural remote sensing and computer vision, and in particular to an automatic identification method for soybean phenological stages based on UAV multimodal data. Background Technology

[0002] Precise monitoring of soybean phenology is crucial for guiding refined farmland management, assessing production potential, and accelerating breeding processes. Traditional phenological monitoring relies primarily on manual field observations, which are highly subjective, time-consuming, and labor-intensive, making them unsuitable for large-scale and high-frequency monitoring needs. While satellite remote sensing technology can achieve large-scale crop phenological monitoring, its spatial resolution is limited, it is significantly affected by cloud and rain conditions, and, more importantly, its analysis typically depends on time-series data throughout the entire growth period, resulting in a significant time lag and failing to meet the needs of dynamic intra-season management and precise decision-making.

[0003] In recent years, unmanned aerial vehicle (UAV) remote sensing has become an important tool for field phenomics and precise crop monitoring due to its high spatiotemporal resolution and flexible operation capabilities. Existing UAV-based phenological identification methods mainly fall into two categories: one is traditional machine learning (ML) methods based on "feature engineering," which extract multi-dimensional features such as vegetation indices, texture features, and structural parameters from images, and then use models such as support vector machines (SVM) and random forests (RF) for classification. This type of method has promoted the development of UAV phenological analysis to some extent, but its performance is highly dependent on manual feature design and selection, and its model generalization ability is limited. The other category is automatic feature learning methods based on deep learning (DL), especially convolutional neural networks (CNNs), which can directly and automatically learn abstract features from images, avoiding tedious feature engineering and providing a new approach for rapid and automated identification of crop phenological stages.

[0004] However, current research still faces several challenges. First, relying solely on image data (deep learning) or tabular feature data (machine learning) has limitations in information representation, making it difficult to fully characterize the complexity of crop phenological changes. Second, most current multimodal data fusion methods remain at the level of simple feature concatenation or linear weighting, lacking a fusion mechanism that can dynamically adjust and intelligently weigh the contributions of different information sources based on data characteristics.

[0005] Therefore, it is evident that existing technologies face the following challenges:

[0006] 1. Limited recognition accuracy due to a single information source. Existing methods rely solely on tabular or image features, which can easily lead to misjudgments when distinguishing consecutive phenological periods with similar morphological features, making it difficult to meet the needs of refined recognition.

[0007] 2. Multimodal data fusion strategies lack intelligence and adaptability. Current fusion methods are mostly based on simple feature concatenation or fixed weighting, which cannot dynamically adjust the importance of different data sources according to sample differences, resulting in low information utilization and insufficient model generalization ability.

[0008] 3. Lack of a technical system that combines high precision, automation, and real-time performance. Traditional manual observation methods are highly subjective and inefficient, while satellite remote sensing suffers from low temporal resolution and weather interference. An end-to-end intelligent framework suitable for UAV platforms for precise monitoring of soybean phenological stages has not yet been formed.

[0009] In view of this, we propose an automatic soybean phenological stage identification method based on UAV multimodal data to solve the existing problems. Summary of the Invention

[0010] The purpose of this invention is to provide an automatic identification method for soybean phenological stages based on UAV multimodal data, so as to solve the problems mentioned in the background art.

[0011] To achieve the above objectives, the present invention provides the following technical solution: an automatic identification method for soybean phenological stages based on UAV multimodal data, comprising the following steps:

[0012] S1: Using a drone platform equipped with an RGB camera and a multispectral camera, high-resolution canopy images were collected in multiple soybean experimental areas;

[0013] S2: The collected UAV images are stitched, registered, and cropped to obtain independent sample images for each cell;

[0014] S3: Using the extracted tabular data, the system evaluates the performance of various machine learning models and their ensemble strategies. Through performance comparison and cross-validation, the optimal machine learning source classifier is determined as the Soft Voting ensemble model.

[0015] S4: Using ConvNeXt as the baseline model, a deep learning model BC-ConvNeXt was constructed. This model introduces a convolutional attention module and a weighted bidirectional feature pyramid network to enhance the ability to extract fine-grained features from soybean canopy images. After comparison with other models, it was determined to be the optimal deep learning source classifier.

[0016] S5: Design a dynamic gated fusion network that takes the predicted probabilities of two optimal source classifiers as input; construct a gated unit composed of a multilayer perceptron that takes the predicted probability vector of the machine learning model as input, learns and generates a dynamic weight g for the current sample, where g is between 0 and 1.

[0017] S6: Dynamic weights g are used to weight and fuse the outputs of the two source classifiers to achieve classification decisions; weights g represent the model's "trust" in the results of the deep learning path, thereby achieving complementary advantages between machine learning and deep learning models in different phenological stages and improving the overall accuracy and stability of recognition.

[0018] Furthermore, during the data acquisition process, S1 synchronously records the acquisition date and corresponding seeding date of each flight mission, providing support for the subsequent construction of time features.

[0019] Furthermore, during the image processing by S2, the extracted image data includes image data and tabular data; among them, the image data is directly used as the input of the deep learning model, and the tabular data includes spectral features, texture features, structural features, and temporal features.

[0020] Furthermore, in S3, machine learning models include logistic regression, K-nearest neighbors, support vector machines, random forests, multilayer perceptrons, XGBoost, LightGBM, and CatBoost, and ensemble strategies include stacking and soft voting.

[0021] Furthermore, in S4, the model input image is processed through a starting layer and four stage layers to extract multi-scale features.

[0022] Furthermore, in S4, each module of the model consists of a 7×7 depthwise separable convolution, a layer normalization, a 1×1 pointwise convolution, a GELU activation function, and another 1×1 pointwise convolution, and is implemented through residual connections.

[0023] Furthermore, in S4, a dual-module enhancement mechanism is introduced at the back end of the backbone network. Multi-scale feature fusion inputs different scale features output from different stages of the backbone network into the weighted bidirectional feature pyramid network, realizing the fusion of deep semantic information and shallow spatial information, and outputting 256-channel fused features.

[0024] Furthermore, in S4, the attention-enhanced fused feature maps are processed through adaptive average pooling, flattening, Dropout regularization, and a fully connected linear layer to output prediction results for seven phenological categories.

[0025] Furthermore, in S5, the network structure consists of an input layer, two hidden layers each containing 128 neurons, and a single-neuron output layer, used to achieve nonlinear feature mapping and weight generation.

[0026] Furthermore, in S5, the two optimal source classifiers include Soft Voting and BC-ConvNeXt.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] 1. The dynamic gating fusion model proposed in this invention achieved an F1 score of 94.3% in the soybean phenological stage identification task, significantly outperforming all single-path models, including the best machine learning model (Soft Voting, F1=92.9%) and deep learning model (BC-ConvNeXt, F1=85.3%). This result demonstrates that this invention achieves a breakthrough improvement in both classification accuracy and stability, exhibiting superior classification precision.

[0029] 2. The dynamic gating network used in this invention no longer relies on traditional fixed weighting or simple averaging, but achieves intelligent arbitration through a data-driven approach, possessing an innovative intelligent fusion mechanism. The model can adaptively learn the information contribution relationship between different sample features, forming a dynamic decision-making strategy of "ML priority, DL assistance": in most cases, it relies on the more robust ML model, and only on specific samples where the ML model may confuse, does it dynamically transfer the decision-making power to the DL model.

[0030] 3. By fusing complementary source classifiers, this invention effectively overcomes the shortcomings of single models in identifying specific phenological stages. The results show that the later stages (R6 and R7) that the ML model struggles to distinguish and the early stages (R1 and R2) that the DL model struggles to distinguish are significantly corrected after fusion, and the number of misclassified samples is greatly reduced, demonstrating a strong collaborative error correction capability and significant collaborative error correction ability.

[0031] 4. The entire framework can automatically classify soybean phenological stages based on a single aerial image taken by a drone, overcoming the time lag problem of traditional satellite remote sensing. The framework has high throughput, low latency and scalability, and can provide near real-time intelligent decision support for field management, significantly improving monitoring efficiency and application value, and has high throughput and real-time monitoring capabilities.

[0032] 5. This invention systematically compares the advantages and limitations of technical routes from machine learning and deep learning to multimodal fusion, providing a reusable algorithm framework and technical reference for subsequent research in crop phenotypic recognition, multimodal information fusion and intelligent agricultural detection, and possesses complete technical verification and scalability. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the workflow of an automatic soybean phenological stage identification method based on UAV multimodal data according to the present invention.

[0034] Figure 2 This is a schematic diagram of the BC-ConvNeXt deep learning model architecture of the present invention;

[0035] Figure 3This is a schematic diagram illustrating the working principle of the dynamic gating fusion network of the present invention;

[0036] Figure 4 This is a comparison chart of confusion matrices for different models of the present invention;

[0037] Figure 5 This is a weight distribution diagram of the dynamic gating model of the present invention. Detailed Implementation

[0038] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. Example 1

[0039] like Figure 1 As shown, the workflow of an automatic soybean phenological stage identification method based on UAV multimodal data includes image data acquisition and preprocessing, source classifier construction and selection, implementation of a dynamic gating fusion model, and classification result output and evaluation. Image data acquisition and preprocessing includes image data collection and ground truth acquisition, image data preprocessing, feature extraction, and dataset construction. Source classifier construction and selection includes building machine learning models based on tabular data, building deep learning models based on image data, and model training. Implementation of the dynamic gating fusion model includes generating outlier prediction data and designing and training the gating network. Classification result output and evaluation include result output and performance verification.

[0040] The specific steps are as follows:

[0041] Image data acquisition and preprocessing:

[0042] 1. Image data acquisition and ground truth acquisition:

[0043] 1) Drone Image Acquisition: Data acquisition was performed using a DJI M300 RTK drone platform, equipped with a DJI Zenmuse P1 (RGB camera) and a MicaSense (multispectral camera). Acquisition was conducted in clear, windless weather to ensure image quality and lighting consistency. Drone flight parameters were set as follows: 30-meter flight altitude, 3.0 m / s flight speed, and 80% forward and lateral overlap. The ground resolutions of the acquired images were 3.69mm (RGB) and 6.67mm (multispectral). To ensure radiometric consistency of the multispectral data, a standard reflectance calibration board was photographed before each flight mission for radiometric calibration and reflectance correction of subsequent images.

[0044] 2) Ground Truth Acquisition: Throughout the soybean growing season, a field survey is conducted every 5-7 days. Agricultural experts manually record the average phenological stage (V1, V2, R1, R2, R6, R7, R8) of each plot as ground truth labels for model training and validation.

[0045] 2. Image data preprocessing:

[0046] 1) Image stitching: The independent images acquired from each flight were stitched together using professional photogrammetry software (Agisoft Metashape) to form an orthophoto map of the entire test area. The output coordinate system adopted was WGS-84.

[0047] 2) Image Registration and Cropping: Using ArcGIS 10.8 software, RGB and multispectral orthophotos were registered with ground control points (GCPs) as the reference to ensure accurate and consistent spatial correspondence among multi-source images. Subsequently, combined with the vector data (shapefile) of the experimental cell boundaries, the orthophotos were automatically batch-cropped to obtain independent images at the cell level.

[0048] 3. Feature extraction and dataset construction:

[0049] 1) Tabular feature extraction:

[0050] Spectral characteristics: Based on multispectral imagery acquired by UAVs, several typical spectral vegetation indices (VIs) were calculated to characterize the spectral reflectance properties of the soybean canopy. This invention selected seven representative vegetation indices, including NDVI, EVI, GNDVI, GRVI, kNDVI, NDRE, and OSAVI.

[0051] Texture feature extraction: Based on the three bands of RGB imagery, the Gray Co-occurrence Matrix (GLCM) method was used to extract eight texture features (mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment, and correlation), resulting in a total of 3×8=24 texture features.

[0052] Structural feature extraction: Using ArcGIS 10.8 software, the average plant height (PH) of each plot was calculated by differentiating between the digital surface model (DSM) and the digital elevation model (DEM) to characterize the population structure.

[0053] Temporal feature extraction: Based on the date of each UAV image acquisition and the date of sowing in the experimental field, the corresponding number of days after sowing (DAS) is calculated to reflect the time progress information of the crop.

[0054] 2) Dataset Construction: All extracted tabular features (spectral, texture, structural, and temporal features) are associated with their corresponding phenological period labels (V1, V2, R1, R2, R6, R7, R8) to form a complete tabular feature dataset. Simultaneously, cropped RGB images of the communities are paired with their corresponding labels to form an image dataset. Subsequently, the comprehensive dataset is divided into training, validation, and test sets (e.g., 7:2:1) to ensure the scientific rigor and representativeness of model training and evaluation. The training set is used for model fitting, the validation set for parameter tuning and structure optimization, and the test set for final performance evaluation and comparative analysis.

[0055] B. Source classifier construction and selection:

[0056] 1. Building a machine learning model based on tabular data:

[0057] 1) Model Benchmarking: Using the tabular dataset constructed in step A, a systematic evaluation of eight mainstream machine learning algorithms was conducted using the 5-fold cross-validation framework. The test models included: Logistic Regression, K-Nearest Neighbors, Support Vector Machine, Random Forest, Multilayer Perceptron, XGBoost, LightGBM, and CatBoost. To ensure a fair comparison of the models' optimal performance, a grid search method was used to automatically optimize the key hyperparameters of each model, and the average F1 score on the validation set was used as the primary performance indicator.

[0058] 2) Ensemble Model Construction and Selection: Based on the benchmark test results of single models, the top four models with the best overall performance in the benchmark test were selected as base learners to construct two ensemble models: a soft voting model and a stacking model. By systematically comparing the performance of the two ensemble methods on the validation set, the soft voting ensemble model was finally determined as the optimal classifier in the machine learning path, as it outperformed both the single model and the stacking model in terms of classification accuracy and stability.

[0059] 2. Construction of deep learning models based on image data, such as... Figure 2 As shown:

[0060] 1) Model architecture design: such as Figure 2 As shown in figure a, ConvNeXtV2 is used as the basic backbone network, and it is modularly optimized and enhanced to construct an improved new model BC-ConvNeXt. The model input is an image of (B, 3, 224, 224), which extracts multi-scale features through one Stem layer and four Stage layers.

[0061] 2) ConvNeXtV core module: such as Figure 2 As shown in b, each module consists of a 7×7 depthwise separable convolution, a layer normalization function (LayerNorm), a 1×1 pointwise convolution (for channel expansion), a GELU activation function, and another 1×1 pointwise convolution (for channel mapping), and is implemented through residual connections (Add).

[0062] 3) Neck attention module, such as Figure 2 As shown in c:

[0063] To enhance the model's feature fusion and attention capabilities, a dual-module enhancement mechanism is introduced at the back end of the backbone network: Multi-Scale Feature Fusion (BiFPN). This mechanism inputs features of different scales from different stages of the backbone network into a weighted bidirectional feature pyramid network (BiFPN Block), achieving efficient fusion of deep semantic information and shallow spatial information, and outputting 256 channels of fused features.

[0064] Convolutional Block Attention Module (CBAM): The BiFPN output features are weighted with both channel attention and spatial attention, enabling the model to adaptively focus on the most discriminative image regions and semantic features, thereby improving the ability to distinguish phenological periods.

[0065] 4) Category header module: such as Figure 2 As shown in d, the fused feature map after attention enhancement is processed by adaptive average pooling, flattening, dropout regularization and fully connected linear layer to output the prediction results of 7 phenological period categories (V1, V2, R1, R2, R6, R7, R8).

[0066] 3. Model Training:

[0067] To fully exploit the expressive power of image features and improve classification accuracy, this invention introduces a two-stage transfer learning strategy during the training of the deep learning source classifier. Stage 1: Backbone Freezing, training only the classification head with a relatively high learning rate (3e-5); Stage 2: End-to-end fine-tuning, unfreezing all network layers and fine-tuning the entire network end-to-end with a smaller learning rate (1e-6). Mixup data augmentation and early stopping mechanisms are employed during training. After training, the model's performance on the validation and test sets is comprehensively evaluated. Using the F1 score as the primary metric, the model with the best performance is selected as the final deep learning source classifier. Ultimately, the BC-ConvNeXt model is determined as the optimal deep learning source classifier.

[0068] C. Implementation of dynamic gating fusion model, such as Figure 3 As shown:

[0069] 1. Out-of-Flight (OOF) data generation:

[0070] To ensure no information leakage during the training of the fusion model and to achieve sample-level cross-validation, this invention employs 5-fold cross-validation to generate out-of-fold prediction data. Specifically, in each fold (fold k) of the 5-fold cross-validation, the training data of that fold (excluding all data from the k-th fold) is used to train both the machine learning source classifier and the deep learning source classifier. The trained models are then used to predict the validation data (the k-th fold data) to obtain the prediction probabilities. This process is repeated 5 times to construct a prediction probability dataset (P) that is perfectly aligned with the true labels and contains no information leakage. ML and P DL This dataset serves as the input data basis for the subsequent dynamic gating fusion model.

[0071] 2. Gated Network Design and Training:

[0072] 1) Gated Unit Structure Design: Construct a gated unit composed of a multilayer perceptron (MLP). Its network structure consists of an input layer, two hidden layers each containing 128 neurons (using the ReLU activation function), and a single-neuron output layer, which is used to realize nonlinear feature mapping and weight generation.

[0073] 2) Dynamic weight generation: The out-of-flight (OOF) prediction probability P of the machine learning source classifier is... ML As input, the gating unit calculates and outputs a scalar value. This output is normalized to the 0-1 interval using the sigmoid activation function to obtain the dynamic weight g. Its calculation formula is: g = σ(MLP(P) ML )).

[0074] 3) Weighted fusion: The predicted probabilities of the two source classifiers (i.e., the machine learning classifier and the deep learning classifier) ​​are weighted and fused using the dynamic weights g output by the gating unit to obtain the final fused prediction probability P. fused The calculation formula is: P fused =g·P DL +(1-g)·P ML .

[0075] 4) Model training:

[0076] The model employs a loss method based on cross-entropy (L... CE The optimization is performed using a composite loss function consisting of the loss function and the balanced regularization loss. Its calculation formula is L. total =LCE (P fused , y true )+λ·mean((g-0.5) 2 The first term is used to measure the prediction probability P. fused With real label y true The first term is the standard cross-entropy loss; the second term is the balance regularization loss, which penalizes extreme weights by calculating the mean squared error between the gated weights g and the center point 0.5; λ is a hyperparameter controlling the strength of regularization. During training, only the parameters of the gated units (MLPs) are updated via backpropagation.

[0077] D. Classification result output and evaluation:

[0078] 1. Output Results: For a new test sample (containing image data and tabular data), the results are first fed into a pre-trained machine learning source classifier and a deep learning source classifier to obtain their respective probabilities. Then, these probabilities are fed into a trained gating unit to generate dynamic weights, and the final fusion probability is calculated according to the fusion formula. The class with the highest probability value is the final output of this invention.

[0079] 2. Performance Validation Method: The performance validation of this invention is achieved by comparing the differences between the model's prediction results and the ground truth labels on the test set. The evaluation process uses commonly used classification performance metrics in the field, including accuracy, precision, recall, F1 score, Matthews correlation coefficient (MCC), and average precision (AP), to comprehensively measure the model's classification accuracy and robustness. Furthermore, as... Figure 4 and Figure 5 As shown, the specific behavior and underlying mechanisms of the model can be analyzed by generating a confusion matrix and analyzing the distribution of gating weights.

[0080] The results of this invention demonstrate that a comprehensive intelligent analysis system for high-precision identification of soybean phenological stages can be constructed based on a machine learning source classifier module (Soft Voting), a deep learning source classifier module (BC-ConvNeXt), and a core dynamic gating fusion module. This system achieves adaptive modeling and accurate discrimination of complex spatiotemporal features by dynamically fusing information from different modalities and algorithmic sources.

[0081] This system can run in cloud or local server environments, efficiently processing and analyzing multi-source phenotypic data collected by drones in batches. After compression and lightweighting, the model can also be deployed on drone-borne computing platforms or agricultural Internet of Things (AIoT) edge computing devices, thereby achieving near real-time intelligent identification and early warning of key soybean growth stages, providing decision support for precision agricultural management.

[0082] The above specific embodiments are merely several preferred embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A method for automatic identification of soybean phenological stages based on UAV multimodal data, characterized in that, The operating steps include: S1: Using a drone platform equipped with an RGB camera and a multispectral camera, high-resolution canopy images were collected in multiple soybean experimental areas; S2: The collected UAV images are stitched, registered, and cropped to obtain independent sample images for each cell; S3: Using the extracted tabular data, the system evaluates the performance of various machine learning models and their ensemble strategies. Through performance comparison and cross-validation, the optimal machine learning source classifier is determined as the Soft Voting ensemble model. S4: Using ConvNeXt as the baseline model, we constructed a deep learning model BC-ConvNeXt. This model introduces a convolutional attention module and a weighted bidirectional feature pyramid network, and after comparison of the models, it was determined to be the optimal deep learning source classifier. S5: Design a dynamic gated fusion network that takes the predicted probabilities of two optimal source classifiers as input; construct a gated unit composed of multilayer perceptrons that takes the predicted probability vector of the machine learning model as input, learns and generates a dynamic weight for the current sample. S6: Dynamic weights are used to weight and fuse the outputs of the two source classifiers to achieve classification decisions; among them, the dynamic weight g represents the model's confidence in the results of the deep learning path.

2. The method for automatic identification of soybean phenological stages based on UAV multimodal data according to claim 1, characterized in that: During the data acquisition process, S1 synchronously records the acquisition date and corresponding seeding date of each flight mission, providing support for the subsequent construction of time features.

3. The method for automatic identification of soybean phenological stages based on UAV multimodal data according to claim 1, characterized in that: During the image processing process, S2 extracts image data including image data and tabular data. The image data is directly used as input to the deep learning model, while the tabular data includes spectral features, texture features, structural features, and temporal features.

4. The method for automatic identification of soybean phenological stages based on UAV multimodal data according to claim 1, characterized in that: In S3, machine learning models include logistic regression, K-nearest neighbors, support vector machine, random forest, multilayer perceptron, XGBoost, LightGBM, and CatBoost, and ensemble strategies include stacking and soft voting.

5. The method for automatic identification of soybean phenological stages based on UAV multimodal data according to claim 1, characterized in that: In S4, the model input image is processed through a starting layer and four stage layers to extract multi-scale features.

6. The method for automatic identification of soybean phenological stages based on UAV multimodal data according to claim 1, characterized in that: In S4, each module of the model consists of a 7×7 depthwise separable convolution, a layer normalization, a 1×1 pointwise convolution, a GELU activation function, and another 1×1 pointwise convolution, and is implemented through residual connections.

7. The method for automatic identification of soybean phenological stages based on UAV multimodal data according to claim 1, characterized in that: In S4, a dual-module enhancement mechanism is introduced at the back end of the backbone network. Multi-scale feature fusion inputs different scale features output from different stages of the backbone network into the weighted bidirectional feature pyramid network, realizing the fusion of deep semantic information and shallow spatial information, and outputting 256-channel fused features.

8. The method for automatic identification of soybean phenological stages based on UAV multimodal data according to claim 1, characterized in that: In S4, the attention-enhanced fused feature maps are processed through adaptive average pooling, flattening, Dropout regularization, and fully connected linear layers to output prediction results for seven phenological categories.

9. The method for automatic identification of soybean phenological stages based on UAV multimodal data according to claim 1, characterized in that: In S5, the network structure consists of an input layer, two hidden layers each containing 128 neurons, and a single-neuron output layer, used to achieve non-linear feature mapping and weight generation.

10. The method for automatic identification of soybean phenological stages based on UAV multimodal data according to claim 1, characterized in that: In S5, the two optimal source classifiers are Soft Voting and BC-ConvNeXt.

Images