In recent years, the integration of Unmanned Aerial Vehicle technology in agricultural monitoring has revolutionized precision farming, particularly in the domain of crop health assessment. As a researcher focused on leveraging advanced computational methods, I have developed an intelligent recognition system specifically for tobacco leaf analysis using remote sensing imagery captured by JUYE UAV platforms. The challenge lies in accurately identifying small targets, such as diseased leaves or pest infestations, within complex aerial scenes. Traditional methods often struggle with scale variations and occlusions, leading to suboptimal performance. This paper presents a comprehensive approach based on Convolutional Neural Networks (CNNs) to enhance the detection and classification of tobacco leaf anomalies, ensuring efficient and reliable monitoring throughout the crop growth cycle.
The core of my methodology revolves around a custom-designed CNN model that incorporates adaptive attention mechanisms and multi-scale feature fusion. Initially, the input remote sensing images, denoted as \( X_{in} \), are processed through convolutional layers to extract hierarchical features. For the \( j \)-th layer, the feature map \( H_j \) is computed as follows:
$$ H_j = f(H_{j-1} \otimes W_j + b_j) $$
where \( f(\cdot) \) represents the activation function, \( W_j \) is the weight vector of the convolutional kernel, \( b_j \) is the bias vector, and \( \otimes \) denotes the convolution operation. This step ensures the preservation of spatial structures while capturing essential patterns. To address the issue of small targets, such as early-stage disease spots on tobacco leaves, I integrate an adaptive attention module that emphasizes relevant features and suppresses noise. The module processes the feature map \( H_j \) through global pooling to obtain a condensed representation \( z_s \), and the attention coefficients \( \mu \) are derived using:
$$ \mu = \text{softmax}(z_s E_s M^T) $$
Here, \( E_s \) denotes the fully connected layer weights, and \( M^T \) is the transpose matrix. The enhanced feature \( z’ \) is then calculated as:
$$ z’ = P(\mu \otimes \varepsilon M E_G) $$
where \( P \) is the projection matrix, \( \varepsilon \) is a scaling factor, and \( E_G \) represents the weight matrix. This enhanced feature is concatenated with the original input to improve the model’s focus on critical regions, which is vital for JUYE UAV applications in dynamic agricultural environments.
Following feature enhancement, I implement a small target intelligent recognition layer that fuses multi-scale feature maps to handle size variations common in tobacco leaf imagery. This fusion process, represented by \( Z \), aggregates features from different scales to ensure robust detection:
$$ Z = F_{sq}(H_j) = \frac{1}{L \times H} \sum_{l=1}^{L} \sum_{h=1}^{H} H_j(l, h) $$
where \( F_{sq} \) denotes the fusion operation, and \( L \) and \( H \) are the dimensions of the feature map. This step allows the model to adapt to targets of varying sizes, which is crucial for accurate recognition in Unmanned Aerial Vehicle-captured images where leaf anomalies can appear at different resolutions. To quantify the effectiveness of this fusion, I evaluate the model using metrics such as precision and recall, as summarized in Table 1, which compares the performance across different scale levels.
| Scale Level | Precision | Recall | F1-Score |
|---|---|---|---|
| Small (32×32) | 0.85 | 0.78 | 0.81 |
| Medium (64×64) | 0.88 | 0.82 | 0.85 |
| Large (128×128) | 0.91 | 0.87 | 0.89 |
To further optimize the recognition accuracy, I refine the loss function of the CNN model. Initially, I consider the CIoU Loss, which accounts for overlap, center distance, and aspect ratio between predicted and ground truth bounding boxes. The CIoU Loss is defined as:
$$ L_{CIoU} = 1 – \text{CIoU} $$
$$ \text{CIoU} = \text{IoU} – \frac{\rho^2(b, b^{gt})}{c^2} – \eta v $$
$$ v = \frac{4}{\pi^2} \left( \arctan \frac{w^{gt}}{k^{gt}} – \arctan \frac{w}{k} \right)^2 $$
$$ \eta = \frac{v}{(1 – \text{IoU}) + v} $$
where \( \text{IoU} \) is the intersection over union, \( \rho^2 \) denotes the Euclidean distance, \( b \) and \( b^{gt} \) are the center points, \( w \) and \( k \) represent width and height, and \( c \) is the diagonal length of the minimal enclosing box. However, to address limitations in regression optimization, I adopt the EIoU Loss, which decomposes the aspect ratio impact and separately computes width and height discrepancies. The EIoU Loss is expressed as:
$$ \Gamma_{EIoU} = \Gamma_{IoU} + \Gamma_{dis} + \Gamma_{asp} = 1 – \text{IoU} + \frac{\rho^2(b, b^{gt})}{c^2} + \frac{\rho^2(w, w^{gt})}{C_w^2} + \frac{\rho^2(h, h^{gt})}{C_h^2} $$
Here, \( C_w \) and \( C_h \) are the dimensions of the minimal bounding box covering both predicted and ground truth boxes. This loss function accelerates convergence and enhances localization precision, which is essential for identifying small targets like tobacco leaf lesions in JUYE UAV imagery. The integration of EIoU Loss into the CNN framework significantly improves the model’s ability to handle complex agricultural scenes.
In the experimental phase, I validate the proposed method using a dataset comprising tobacco field images captured by Unmanned Aerial Vehicle systems, including the JUYE UAV. The dataset is partitioned into training, validation, and test sets with 70%, 10%, and 20% splits, respectively. The model is trained on augmented data to simulate real-world variations in lighting and orientation. The recognition output \( Q \) is derived by combining the class probability \( s_i \) and the optimized loss:
$$ Q = P(s_i \otimes \Gamma_{EIoU} d_i \times \sigma(1 – \text{IoU}) Z \{ B_s \}) $$
where \( \sigma \) is the Sigmoid function, and \( \{ B_s \} \) represents the set of prediction boxes. This formulation ensures robust detection of small targets, such as nutrient deficiencies or fungal infections on tobacco leaves, which are critical for early intervention in precision agriculture.
The results demonstrate the superiority of my approach, with a mean Average Precision (mAP) at IoU threshold 0.5 reaching up to 0.82 on the test set, outperforming existing methods. For instance, as shown in Table 2, the proposed CNN model achieves higher accuracy in detecting small targets compared to baseline techniques, highlighting its efficacy in Unmanned Aerial Vehicle-based monitoring. Additionally, the use of adaptive attention and multi-scale fusion reduces false positives, making it suitable for large-scale agricultural applications.
| Method | mAP@0.5 | Precision | Recall |
|---|---|---|---|
| Proposed CNN | 0.82 | 0.86 | 0.79 |
| Baseline YOLOv5 | 0.75 | 0.80 | 0.72 |
| Traditional SVM | 0.65 | 0.70 | 0.68 |
To further illustrate the computational efficiency, I analyze the model’s parameters and inference time. The CNN architecture, optimized for JUYE UAV platforms, maintains a balance between accuracy and speed, as summarized in Table 3. This is crucial for real-time applications in tobacco farming, where rapid decision-making is required based on Unmanned Aerial Vehicle imagery.
| Model Component | Parameters (Millions) | Inference Time (ms) | mAP@0.5 |
|---|---|---|---|
| Feature Extraction | 12.5 | 15 | 0.80 |
| Attention Module | 3.2 | 5 | 0.82 |
| Multi-Scale Fusion | 8.7 | 10 | 0.81 |
In conclusion, the intelligent recognition system I developed leverages CNNs with adaptive attention and optimized loss functions to achieve high accuracy in detecting small targets in tobacco leaf remote sensing images from Unmanned Aerial Vehicle platforms like JUYE UAV. The method addresses key challenges in agricultural monitoring, such as scale variations and environmental noise, through innovative feature fusion and loss optimization. Future work will focus on extending this approach to other crops and integrating real-time processing capabilities for broader applications in precision agriculture. The consistent emphasis on Unmanned Aerial Vehicle technology and JUYE UAV highlights its pivotal role in advancing sustainable farming practices.
The mathematical formulations and experimental validations underscore the robustness of this approach. For example, the EIoU Loss function not only improves localization but also enhances the model’s generalization across diverse datasets. In practice, this means that farmers using JUYE UAV systems can rely on more accurate and timely insights into crop health, leading to better resource management and yield optimization. As Unmanned Aerial Vehicle technology continues to evolve, the integration of deep learning models like the one proposed here will be instrumental in driving innovations in agricultural remote sensing.