Detection of Pine Wilt Disease-Infected Dead Trees in Complex Mountainous Areas Using Enhanced YOLOv5 and UAV Remote Sensing

Abstract

Pine wilt disease endangers the ecological stability of China’s coniferous woodlands. In a specific region, the number of dead pine trees has exhibited a consistent year-on-year increase, highlighting the urgent need for efficient and sustainable monitoring strategies. However, UAV-based remote sensing methods currently face challenges in complex environments, including insufficient feature-capture capabilities, interference from visually similar objects, and limited localization accuracy. This study developed a remote sensing workflow leveraging high-resolution UAV imagery to oversee pine trees affected with pine wilt disease. An enhanced YOLOv5 detection model was employed to identify symptomatic trees. To strengthen feature extraction capabilities—particularly for color and texture traits indicative of infection—different types of attention mechanisms, for instance SE, CBAM, ECA, and CA, were integrated as part of the model. Furthermore, a BiFPN structure was incorporated to enhance the fusion of features across multiple scales, and the EIoU loss function was adopted to boost the accuracy of bounding box prediction, ultimately enhancing detection precision. Experimental results show that the enhanced SEBiE-YOLOv5 framework achieved a precision of 89.4%, with an AP of 86.1% and an F1-score of 83.1%. UAV-based monitoring conducted during the spring and autumn of 2023 identified 616 dead trees, with field verification accuracy ranging from 88.91% to 92.42% and localization errors within 1–10 m. These findings validate the method’s high accuracy and spatial precision in complex mountainous forest environments. By integrating attention mechanisms, BiFPN, and the EIoU loss function, the proposed SEBiE-YOLOv5 model substantially enhances the recognition accuracy of key features in infected trees as well as their localization performance, and offers a practical and computationally efficient approach for the long-term surveillance of pine wilt disease in challenging terrain.

1. Introduction

PWD ranks among the most destructive forest afflictions. Since its introduction in 1982, it has caused significant ecological and economic losses in China. Effective and precise surveillance of trees infected by PWD is crucial to prevent its transmission. Epidemiological surveys serve as the foundation for epidemic prevention and control. Currently, primary monitoring approaches include ground surveys, satellite remote sensing, and UAV remote sensing. However, traditional ground surveys are constrained by challenging terrain and limited spatial coverage, reducing their effectiveness in large-scale forest monitoring and thereby facilitating the spread of the disease. With the advancement of remote sensing technology, satellite-based approaches have been widely applied in forestry pest and disease detection due to their broad coverage and capability for periodic observations [1]. For instance, Lin et al. [2] extracted coniferous forest information from Landsat imagery and achieved a classification accuracy of 81.67% using the random forest algorithm. Hao et al. [3] utilized MODIS remote sensing imagery to analyze the occurrence and spread of pine wilt disease, providing a scientific basis for its long-term management and control in the Yangtze River Basin. Huang et al. [4] employed high-resolution Gaofen-1 and Gaofen-2 imagery combined with deep learning to detect infected trees with high accuracy. Although these satellite remote sensing techniques offer distinct advantages for large-scale monitoring of pine wilt disease and infected trees, limitations remain in achieving precise detection. Medium- and low-resolution imagery (e.g., Landsat and MODIS) enables large-scale monitoring but lacks the spatial resolution to identify individual or small clusters of infected trees. High-spatial-resolution imagery (e.g., GF series, QuickBird, and WorldView) improves spatial detail but is restricted by limited spectral bands, making it difficult to capture the fine spectral characteristics of infected trees. Furthermore, satellite remote sensing, in general, suffers from long revisit cycles and susceptibility to weather conditions, which hinder the acquisition of high-quality imagery. In contrast, recent progress in UAV remote sensing provides high spatial resolution and strong operational flexibility, offering a promising alternative for efficient monitoring of forest pests and diseases [5,6].

Numerous research projects have explored the use of UAV-based methods for identifying trees infected by PWD. As computer technology has advanced, the use of aerial imagery captured by drones combined with machine learning methods has become more common in tracking and identifying forest pests and diseases. For instance, Tao et al. [7] applied HSV thresholding to UAV remote sensing images to identify infected trees, while Liu et al. [8] utilized multi-template matching to detect trees at various stages of infection. Syifa et al. [9] employed Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs) to classify UAV imagery, effectively distinguishing PWD-infected trees from other vegetation types. In another study, Iordache et al. [10] collected both multispectral and hyperspectral data of PWD-affected areas and compared their classification performance using Random Forest (RF) algorithms. Despite these advancements, traditional machine learning methods largely depend on manually engineered features, limiting their ability to effectively extract discriminative information in complex forest environments [11]. In contrast, the rise of deep learning has facilitated automated feature extraction from remote sensing imagery, greatly enhancing the accuracy and reliability of infected tree identification [12,13]. Yu et al. [14] evaluated the effectiveness of two deep learning-powered object detection algorithms, Faster R-CNN (Region-based Convolutional Neural Network) and YOLOv4 (You Only Look Once, version 4), against conventional machine learning techniques such as RF and SVM for identifying trees damaged by PWD. Deng et al. [5] trained a dead pine tree detection model using the Faster R-CNN framework, incorporating a Region Proposal Network (RPN) and Residual Network (ResNet), and further optimized the model’s loss function and anchor settings to enhance detection precision and overall effectiveness. Sun et al. [15] improved model accuracy by replacing the VGG16 backbone with ResNet50 and using K-means clustering for anchor box generation, eliminating the need for manual configuration. Qin et al. [16] proposed the SCANet network, which integrates a spatial information module to retain critical PWD-related features, achieving an identification accuracy of 79%. Hu et al. [17] adopted Mask R-CNN—a multi-task detection framework—as the base model and introduced an improved multi-scale perception block module to extract finer features and reduce missed detections. Additionally, Zhou et al. [18] utilized multispectral imagery to construct a multi-band fusion-based PWD detection model using the YOLOv5 architecture, significantly improving early-stage detection accuracy and thereby strengthening early warning capabilities. Despite the progress achieved in forest stands and small-scale plots, the integration of UAVs with deep learning in practical applications still faces two key challenges: (i) pronounced scene limitations, as most studies have focused on small-scale sample sites or managed forest plantations, while complex mountainous terrains tend to cause high omission rates; (ii) in mountainous environments, topographic undulations and image distortions constrain localization accuracy, reducing the timeliness of infected tree removal and indirectly increasing the risk of disease spread.

Therefore, given the rapid spread of pine wilt disease and the urgent need for timely removal of infected trees [19], as well as the challenges of remote sensing identification in complex mountainous forest environments—such as dense backgrounds and the subtle visual characteristics of infected trees—this research introduces a drone-assisted remote sensing method to identify pine wilt disease-infected dead trees. The workflow integrates high-resolution UAV imagery with the YOLOv5 object detection model, enhanced by multiple visual perception modules. These include attention mechanisms (SE, CBAM, ECA, and CA), a Weighted Bidirectional Feature Pyramid Network (BiFPN), and the EIoU loss function for optimized bounding box regression. By constructing and comparing multiple model configurations, the study evaluates the accuracy, robustness, and practical applicability of the proposed method under real-world mountainous forest conditions. The findings show that the system provides live tracking with very little deviation in pinpointing location, delivering an effective and dependable technological approach for managing forest pests and safeguarding resources. This not only helps maintain ecological balance but also supports long-term sustainable forest stewardship.

2. Materials

2.1. Study Area

The research zone is located in Xishan District, Kunming City, Yunnan Province, China (25°03′N–25°08′N, 102°50′E–102°55′E), situated on the northwestern shore of Dian Lake and forming part of the Central Yunnan Plateau. The terrain is markedly undulating, with elevations descending from northwest to southeast. The average elevation is 2032 m, with the highest point reaching 2803 m. The area has a subtropical semi-humid monsoon climate, averaging 16–18 °C annually with roughly 1094.1 mm of yearly rainfall. Xishan District is rich in forest resources, exhibiting a forest coverage rate of 53.87%, and is predominantly covered by mixed forests. Due to the coexistence of suitable host plants (Pinus yunnanensis Franch) and host insects (Monochamus alternatus), as well as favorable environmental conditions, PWD has spread extensively throughout the area. In recognition of its severity, Xishan District was designated as a key epidemic zone in Announcement No. 6 of 2022 by the National Forestry and Grassland Administration. The National Forest Resources Subcompartment Survey—also referred to as the Forest Resources Planning and Design Survey—is a systematic investigation that uses state-owned forest farms, nature reserves, forest parks, or county-level administrative regions as the basic survey units, aiming to provide data support for the development of forest management plans, master designs, and forestry zoning and planning. Based on the data from the fourth Forest Resources Subcompartment Survey of Yunnan Province, 95 pine-dominated forest plots within the study area were selected to monitor tree mortality caused by PWD. The primary tree species in these plots are Yunnan Pine and Armand Pine. Figure 1 displays the study region along with the locations of the monitoring plots.

2.2. Data Acquisition

Field data were obtained with a DJI Phantom 4 RTK UAV fitted with an RGB camera (8.8 mm focal length). From July to August 2022, typical pine forest areas within the study region containing pine wilt disease-induced dead pine trees were selected, and a total of 9526 images with a resolution of 5472 × 3648 pixels (3:2 aspect ratio) were acquired. Drawing on the research findings of Ma et al. [20] and Wang et al. [21], the data acquisition altitude in this study was set to 50–100 m. Two additional UAV campaigns were conducted over the monitored plots in spring and autumn of 2023, respectively, each producing 1073 images. Flight trajectories were planned using monitoring plot boundaries and 12.5 m-resolution ALOS DEM data. A terrain-following flight mode was adopted at an average altitude of 70 m, calculated using a self-developed path-planning algorithm. Experimental validation confirmed that this altitude ensured sufficient image resolution for detecting PWD-affected trees, capturing distinct visual features while maintaining an image overlap rate below 15%. This configuration achieved a balance between image quality, spatial coverage, and UAV flight endurance, enabling efficient large-scale data acquisition. An example of the UAV data collection setup is shown in Figure 2.

2.3. Data Preprocessing and Dataset Development

2.3.1. Data Preprocessing

The 9526 UAV-acquired RGB images of pine trees infected with PWD were subjected to a multi-stage preprocessing workflow. An initial background filtering step was employed to eliminate irrelevant samples and improve dataset quality. To improve the detection model’s ability to generalize, the filtered images were augmented with horizontal flips, vertical flips, and rotations of 90° and 180°. To support real-time monitoring applications, the original image dimensions were retained, allowing UAV-acquired images to be directly input into the model without the need for secondary cropping. Following this process, a final dataset comprising 2760 images was generated, each at 5472 × 3648 pixels with a spatial resolution of 72 dpi in both horizontal and vertical directions. These images clearly display distinguishing visual features of PWD-affected trees, providing a robust foundation for subsequent image recognition, analysis, and the training and evaluation of deep learning models.

2.3.2. Annotation and Dataset Partitioning

As one of the primary distribution areas of Yunnan Pine, Xishan District in Kunming City was designated a key epidemic zone for pine wilt disease in 2022 by the Chinese government. Based on field surveys combined with UAV remote sensing, the typical pathological features of infected trees were identified: the crowns of diseased trees exhibit yellow, reddish-brown, or whitish-brown discoloration and mortality, forming pronounced spectral–morphological contrasts with healthy pines. Meanwhile, two categories of high-interference targets were observed in the study area: reddish-brown broadleaf trees and whitish-brown non-PWD dead trees, both of which display texture and color characteristics highly similar to those of infected pines. Given that the primary objective of this study was to screen all potential sources of dead pine trees across the entire region, the dataset was divided into two categories. The first category, “sick_tree,” includes PWD-infected dead pines (yellow, reddish-brown, and whitish-brown infected trees). The second category, “sick_other,” comprises non-PWD trees with similar appearances, including reddish-brown broadleaf trees and other dead trees with colors resembling infected pines. Both categories were used for model training and testing to enhance the model’s capability to accurately distinguish between them. Example dataset images are presented in

Table 1. Dataset Sample Images.

Type of Labeling	Sick_Tree			Sick_Other
style	Yellow-stage pine	Reddish-brown-stage pine	White-brown-stage diseased dead pine	Red broadleaf trees	Other dead trees
Canopy

.

To construct the target detection model, PWD-infected trees were annotated using the LabelImg software. Each target instance was enclosed within a rectangular bounding box, with the label sick_tree assigned to dead pine trees and sick_other to the “Other Trees” class. The annotations were stored in VOC format (XML files). After completing the annotation process, the dataset was randomly divided into training and testing subsets using a 9:1 ratio, which produced 2484 training images and 276 testing images.

3. Methods

3.1. Conceptual Framework and Methodology

Pine trees infected with PWD typically die within approximately 40 days after the onset of symptoms. As such, rapid and accurate identification of dead trees, followed by timely removal, is critical for effectively containing the spread of the disease. For real-time surveillance and prompt intervention, the research utilizes YOLOv5 as the fundamental detection architecture. YOLOv5 was selected for its robust real-time performance, high detection accuracy, demonstrated effectiveness in capturing visual features of PWD-infected trees, and compatibility with deployment on mobile and edge devices. The overall technical workflow of this study is as follows:

(1)

Data Collection and Screening: UAVs with visible-light cameras captured images of dead pine trees infected by PWD in Xishan District. A manual screening process was conducted to ensure data quality; images that were blurred or lacked clear visual features of infection were excluded. A sample library of dead infected trees was constructed by retaining high-quality imagery that clearly depicted the characteristics of infected trees.

(2)

Data Preprocessing and Dataset Partitioning: The screened images were augmented through horizontal flipping, vertical flipping, and 90° and 180° rotations to enhance model generalization. Annotation of PWD-infected trees was completed via the LabelImg tool, with bounding boxes used to mark each target instance. The annotated dataset was subsequently split into training and testing subsets at a 9:1 ratio, supplying data for model training and performance evaluation.

(3)

Optimal Model Selection: This study leveraged the comprehensive advantages of the YOLO series algorithms in terms of accuracy, speed, and real-time performance, selecting YOLOv5 and YOLOv8 for comparative experiments. The results demonstrated that YOLOv5 achieved a superior balance between accuracy and efficiency, and was therefore adopted as the baseline model. Its lightweight architecture—characterized by small model size and fast inference speed—enables deployment on mobile platforms and supports real-time monitoring scenarios, thereby facilitating precise localization of infected trees and informed decision-making for their removal. To address the challenges of insufficient feature extraction and interference from spectrally similar objects in complex environments, targeted improvements to the baseline model were proposed as follows:

Feature extraction of PWD-infected trees in complex mountainous terrain presents considerable challenges. Additionally, previously undetected dead trees from earlier outbreaks continue to contribute significantly to the ongoing spread of the disease [22,23]. To address these challenges, attention mechanisms were introduced to enhance the network’s representational power. These mechanisms enhance the model’s capacity to extract critical features—particularly the texture and color of infected trees—by highlighting pertinent information and diminishing background noise, thus lowering the missed detection rate. Specifically, the basic BottleNeck module within the C3 component of the YOLOv5 backbone was replaced with four kinds of attention modules: Squeeze-and-Excitation (SE) [24], Convolutional Block Attention Module (CBAM) [25], Efficient Channel Attention (ECA) [26], and Coordinate Attention (CA). These modules introduce dynamic weight allocation mechanisms across feature channels, thereby enhancing the extraction of discriminative features related to infected tree characteristics. The architectural structures of the four attention mechanisms are illustrated in Figure 3a–d.

Visual interference from background elements with similar appearance often results in identification errors, particularly in complex forest environments. While attention mechanisms strengthen the model’s emphasis on crucial features in the backbone network, deeper convolutional layers tend to preserve features of large objects while diminishing those of smaller ones—such as distant or partially occluded infected trees—during feature fusion. To address this limitation, the BiFPN was introduced. BiFPN facilitates efficient bi-directional information flow across multiple feature scales and layers, significantly improving multi-scale feature fusion. Notably, BiFPN incorporates a learnable weighted fusion strategy that assigns adaptive weights to different input features, enabling the network to learn and emphasize the most informative feature maps. This mechanism enhances the integrity of feature representations, mitigates the loss of critical information during convolution, and ultimately improves detection accuracy [27]. Therefore, this study incorporates BiFPN to enhance multi-scale feature fusion and boost the model’s target feature representation. The architecture of the BiFPN module is depicted in Figure 3e.

For object detection, the loss metric measures the variance between forecasted and actual bounding boxes, thereby directing model refinement throughout the training phase. YOLOv5 adopts the CIoU loss function [28], integrating elements like overlap region, centroid distance, and shape ratio. However, CIoU’s aspect ratio term lacks precision in differentiating width and height variations, which can hinder optimization effectiveness [29]. To address this limitation, the EIoU loss function [30] is introduced in this study as a replacement. EIoU, distinct from CIoU, individually calculates width and height discrepancies, enhancing both convergence velocity and regression precision. Furthermore, the EIoU technique enhances sample balance by concentrating the model’s attention on premium anchor boxes more efficiently. This improves object detection performance.

Based on the aforementioned enhancement strategies, four improved detection models were developed: SEBiE-YOLOv5, CBAMBiE-YOLOv5, ECABiE-YOLOv5, and CABiE-YOLOv5. These models retain YOLOv5′s advantages in mobile deployment and real-time performance, while incorporating attention mechanisms, BiFPN, and EIoU to improve detection accuracy and robustness. To assess their performance in detecting PWD-infected dead trees, comparative tests were run using the prepared training and test datasets. The detection performance of each model was analyzed to identify the most suitable architecture for practical applications in large-scale monitoring of forest health.

(4)

Dynamic Application and Monitoring: The optimal model identified through comparative analysis was deployed to process UAV imagery collected in Xishan District during the spring and autumn of 2023. The model was exported in TorchScript format for efficient deployment. UAV flight paths were planned based on the boundaries of the monitored forest plots, and the acquired images were directly input into the detection model for automated identification of PWD-infected dead trees. Upon detection, bounding boxes were generated around infected targets, and the pixel coordinates of the bounding box centers were extracted. Using the georeferenced coordinates (latitude and longitude) of the image centers and known spatial resolution parameters, the pixel coordinates were converted into geographic coordinates. These geographic locations were then associated with the corresponding detected trees and outputted to enable spatial localization. The resulting coordinate data provided a basis for subsequent field verification and on-site disposal of infected trees.

The general process of this study is depicted in Figure 3.

3.2. Experimental Environment and Evaluation Indicators

All experiments were performed on a Windows 11 64-bit platform. The hardware and software configurations are summarized in

Table 2. Configuration table of hardware and software.

Software and Hardware	Technical Parameters
CPU	13th Gen Intel(R) Core(TM) i7-13650HX
GPU	NVIDIA GeForce RTX 4060 Laptop GPU
Memory/GB	16
Storage Capacity/TB	1
Deep Learning Framework	Pytorch 2.4.1
Programming Language	Python
Marking Software	LabelImg 1.8.6

. The system was equipped with a 13th Gen Intel (R) Core™ i7-13650HX CPU and an NVIDIA GeForce RTX 4060 Laptop GPU with 16 GB of VRAM, running CUDA version 12.4. The deep learning framework was implemented in Python 3.8. The model was trained using a starting learning rate of 0.01 across 300 predefined epochs. With a batch size of 8, each epoch included 130–250 training steps. Upon model convergence, the optimal weight parameters were saved for subsequent inference and prediction tasks.

This research employed four key indicators to gauge model accuracy: precision (P), recall (R), mean average precision ($AP$), and the F1 score (F1). To measure processing efficiency, we tracked frames per second (FPS), which serves as a reliable benchmark for evaluating each model’s real-time detection capabilities.

The formula used to calculate the above quantitative evaluation metrics is as follows:

P denotes the percentage of correctly identified true positives among predicted positives:

$$P={\displaystyle \frac{TP}{TP+FP}}$$

(1)

where TP and FP represent true positives and false positives, respectively.

R represents the true positive rate, indicating the percentage of genuine positive cases accurately detected by the model:

$$R={\displaystyle \frac{TP}{TP+FN}}$$

(2)

where FN denotes false negatives.

$AP$ assesses how well the model performs in detection by measuring the accuracy of predicted bounding boxes and the model’s ability to correctly identify dead pine trees infected with PWD:

$$AP={\displaystyle {\int }_0^{1}P(R)dR}$$

(3)

where $P(R)$ denotes precision at recall level R.

F1-score represents the harmonic mean of precision and recall, providing a comprehensive measure of both metrics to evaluate the overall performance of the model in object detection:

$$F1={\displaystyle \frac{2\times P\times R}{P+R}}$$

(4)

where P denotes precision and R represents recall (all-rate check).

$FPS$ measures the processing speed of the model, with higher values indicating faster recognition and analysis of imagery, directly reflecting the model’s inference efficiency and its suitability for real-time detection tasks:

$$FPS={\displaystyle \frac{1}{\mathrm{Pr}e-process+Inference+NMS}}$$

(5)

where $\mathrm{Pr}e-process$ is the time required for data preprocessing, $Inference$ is the model inference time, and $NMS$ is the time consumed for optimal detection frame selection.

4. Results

4.1. Comparative Performance Analysis and Validation of Model Enhancements

In this study, the detection performance of YOLOv5 and YOLOv8 baseline models for PWD-infected trees was comparatively analyzed (Figure 4). Results showed that, in infected tree identification, YOLOv5 achieved P, R, AP, and F1 values of 82.0%, 79.4%, 83.4%, and 80.7%, respectively, which were markedly higher than the corresponding values of YOLOv8 (74.6%, 79.4%, 77.3%, and 72.2%). For the identification of other tree species, YOLOv5 also outperformed YOLOv8, with P, R, AP, and F1 values of 36.6%, 59.1%, 36.2%, and 45.4%, compared to 25.5%, 58.6%, 23.1%, and 27.0%, respectively. Although YOLOv8 exhibited a higher detection speed (50.8 FPS) compared to YOLOv5 (34.1 FPS), its high-speed architecture partially compromised its feature extraction capability, resulting in reduced accuracy in infected tree detection. Considering precision, recall, AP, and FPS comprehensively, YOLOv5 was ultimately selected as the baseline model for subsequent improvements in this research.

To further enhance model performance, this study incorporated the BiFPN multi-scale feature fusion network and the EIoU loss function into the YOLOv5 framework, combined with four attention mechanisms—SE, ECA, CBAM, and CA—to develop four improved models: SEBiE-YOLOv5, CBAMBiE-YOLOv5, ECABiE-YOLOv5, and CABiE-YOLOv5. As shown in Figure 5, SEBiE-YOLOv5 achieved the highest performance in detecting infected trees, with a precision of 89.4%, recall of 77.7%, AP of 86.1%, and F1-score of 83.1%, outperforming the other improved models and improving over the original YOLOv5 by 7.7%, 2.7%, and 2.4% in precision, AP, and F1-score, respectively. For other tree species, SEBiE-YOLOv5 reached a precision of 50.3%, a recall of 47.6%, an AP of 44.2%, and an F1-score of 49.0%, surpassing both the other improved models and YOLOv5 by 13.7%, 8.0%, and 3.6% in precision, AP, and F1-score. In terms of inference speed, SEBiE-YOLOv5 achieved 32.7 FPS, slightly lower than YOLOv5 (34.1 FPS) but faster than the other improved models, demonstrating an optimal balance between detection accuracy and computational efficiency. Considering both F1-score and FPS, SEBiE-YOLOv5 exhibited the best overall performance, confirming its effectiveness and suitability for infected tree detection.

To evaluate the detection performance of the proposed models, CABiE-YOLOv5, CBAMBiE-YOLOv5, ECABiE-YOLOv5, SEBiE-YOLOv5, YOLOv5, and YOLOv8 were tested on sick_tree, sick_other, and background detection. As shown in Figure 6, CABiE-YOLOv5 achieved the highest recognition rate of 0.84 for sick_tree, while CBAMBiE-YOLOv5 showed a slightly lower rate of 0.83 with similar performance in other categories. ECABiE-YOLOv5 performed best for sick_other at 0.64, but its sick_tree and background detection decreased to 0.80 and 0.24, respectively. SEBiE-YOLOv5 maintained high sick_tree accuracy at 0.84, achieved 0.62 for sick_other, and 0.31 for background, showing the most balanced performance across the three categories. The original YOLOv5 showed moderate detection across categories, whereas YOLOv8 exhibited the lowest accuracy with 0.68 for sick_tree, 0.26 for sick_other, and 0.16 for background. These results indicate that YOLOv5 models enhanced with attention mechanisms and BiFPN outperform the original YOLOv5 and YOLOv8, with SEBiE-YOLOv5 achieving the optimal overall performance.

4.2. Ablation Experiment

To assess the efficacy of each proposed enhancement, an ablation study was conducted with the original dataset and the optimal model configuration identified in Section 4.1. The YOLOv5 architecture was used as the baseline network. Each of the three modules—SE, BiFPN, and the EIoU loss function—was individually integrated into the baseline model, while keeping all other parameters unchanged, to assess their respective impacts on model performance. The training settings and hyperparameters remained consistent with those used in previous experiments. The inclusion of each module is indicated by “√”, and its exclusion by “-”. The findings from the experiment are presented in

Table 3. Results of Ablation Study.

YOLOv5 Baseline Network	SE	BiFPN	EIoU	P/%	R%	F1%	AP%	FPS
√	-	-	-	82	79.4	80.7	83.4	34.1
√	√	-	-	86.1	77.5	81.7	84.1	32.6
√	-	√	-	87.3	71.6	78.5	81.1	31.9
√	-	-	√	76.7	83.1	79.8	82.6	34.2
√	√	√	-	89	77.7	83	83	31.3
√	-	√	√	85.6	80.2	82.8	82.3	33.3
√	√	-	√	81.7	76.9	79.2	81.5	32.3
√	√	√	√	89.4	77.7	83.1	86.1	32.7

.

The results of the ablation study for the SEBiE-YOLOv5 model are presented in

Table 4. Heatmap Image Comparisons.

	①	②	③	④
Original Image
YOLOv5
SEBiE-YOLOv5
CBAMBiE-YOLOv5
ECABiE-YOLOv5
CABiE-YOLOv5

. Each integrated module demonstrated distinct effects on model performance. The SE module, by recalibrating channel-wise features, improved precision (P) by 4.1% but resulted in a 1.9% decrease in recall (R). The BiFPN module enhanced precision by 5.3% through improved multi-scale feature fusion; however, it also introduced additional network complexity, leading to a 2.2 FPS reduction. The EIoU loss function optimized bounding box regression, yielding a 3.5% increase in recall while maintaining the FPS nearly unchanged. When all three modules were combined, the model achieved the best overall performance. For the identification of dead infected trees (sick_tree), the precision reached 89.4%, the recall was 77.7%, the F1-score improved to 83.1%, the average precision (AP) increased to 86.1%, and the detection speed reached 32.7 FPS. For the identification of other infected trees (sick_other), the model reached a P of 50.3%, R of 47.6%, F1 of 49.0%, and AP of 44.2%. Overall, the detection performance was substantially improved. The findings confirm the efficacy of the suggested improvement tactics. The synergistic integration of the SE, BiFPN, and EIoU modules significantly boosted detection accuracy and overall performance while preserving computational efficiency, thereby offering strong support for YOLOv5-based object detection in complex environments.

4.3. Analysis of Detection Performance and Field Validation

As demonstrated in Section 4.1 and Section 4.2, the YOLOv5 baseline model and its four improved variants—SEBiE-YOLOv5, CBAMBiE-YOLOv5, ECABiE-YOLOv5, and CABiE-YOLOv5—were evaluated for their performance in detecting infected trees in complex scenarios through detection error analysis and heatmap experiments.

Table 5. Detection Results Presentation.

	Scenario a	Scenario b	Scenario c	Scenario d
Original Image
YOLOv5
SEBiE-YOLOv5
CBAMBiE-YOLOv5
ECABiE-YOLOv5
CABiE-YOLOv5

presents the detection results of the baseline and improved models under different background complexities. In simple scenarios (Scenario a), all models achieved a 100% recognition rate for infected trees. In Scenario b, YOLOv5 misidentified red broadleaf trees as PWD-killed trees, while the four improved models correctly classified them as healthy. SEBiE-YOLOv5 reached a confidence level of 90%, higher than YOLOv5 at 81%. In Scenario c, SEBiE-YOLOv5 correctly detected the target without misclassifying healthy trees, whereas CBAMBiE-YOLOv5 and ECABiE-YOLOv5 misclassified healthy trees and red soil into the sick_other category, respectively. CABiE-YOLOv5 detected the target but with a higher false detection rate compared to SEBiE-YOLOv5. In Scenario d, containing a complex background with red soil, pine forests, and bamboo forests, three infected trees were present. SEBiE-YOLOv5 correctly identified all three, while the other models detected only two infected trees, missing one.

To further evaluate model performance, heatmaps were generated for three representative sample images, as shown in Table 5. The YOLOv5 heatmaps exhibited dispersed high-confidence areas in the target region with blurred object boundaries. In contrast, SEBiE-YOLOv5 produced stronger and more concentrated high-confidence responses in the target region with clearer boundaries.

Based on the above detection error analysis and heatmap visualization results, the proposed SEBiE-YOLOv5 model, by integrating the SE attention mechanism, the BiFPN multi-scale feature fusion network, and the EIoU loss function, achieved synergistic optimization of key techniques: it significantly enhanced the perception of color features of infected trees and effectively suppressed interference from complex backgrounds such as vegetation shadows and soil, thereby providing a more robust technical solution for the precise identification of PWD-infected trees.

To assess practical applicability, the model was applied to continuous UAV-based monitoring of pine forests in spring and autumn 2023. Geographic coordinates of detected dead trees were extracted and verified in the field using GPS devices.

Table 6. Monitoring Results Statistics.

Monitoring Period	Number of Dead Trees Monitored (Plants)	Number of Dead Pine Trees Field Verified (Plants)	Number of Other Tree Species Killed (Plants)	Tree Species Accuracy (%)
Spring 2023	550	489	61	88.91
Fall 2023	66	61	5	92.42

summarizes the monitoring results. In spring 2023, 550 dead trees were detected, of which 489 were confirmed as dead pine trees and 61 as other species, yielding a verification accuracy of 88.91%. In autumn 2023, 66 dead trees were detected, with 61 confirmed as dead pine trees and 5 as other species, resulting in 92.42% accuracy.

Table 7. UAV Monitoring and Field Location Information Comparison.

Dead Tree Number	UAV Monitoring of Coordinates		Field Location Coordinates		Position Deviation (m)
	Longitudes	Latitude	Longitudes	Latitude
1	102°35′53.13000″	25°01′46.80000″	102°35′53.06778″	25°01′46.88202″	3.08
2	102°35′57.08000″	25°01′45.99000″	102°35′57.14614″	25°01′45.89928″	3.36
3	102°35′54.20000″	25°01′49.99000″	102°35′54.24086″	25°01′49.88282″	3.51
4	102°36′14.14000″	25°00′55.87000″	102°36′14.12237″	25°00′55.79849″	2.27
5	102°36′13.34000″	25°00′57.18000″	102°36′13.17424″	25°00′57.35415″	7.11
6	102°36′39.63000″	25°02′25.43000″	102°36′39.78296″	25°02′25.59917″	6.76
7	102°36′40.62000″	25°02′25.75000″	102°36′40.79011″	25°02′25.81068″	5.12
8	102°36′23.31000″	25°04′27.56000″	102°36′23.2465″	25°04′27.48446″	2.94
9	102°36′23.20000″	25°04′28.64000″	102°36′23.1797″	25°04′28.4707″	5.27
10	102°36′22.32000″	25°04′27.24000″	102°36′22.52544″	25°04′27.02691″	8.75

provides a comparison between UAV-detected coordinates and field-measured positions for a randomly selected sample, showing positional deviations ranging from 1 to 10 m.

The spatial distribution of PWD-infected trees detected during the two monitoring periods is illustrated in Figure 7. A comparative analysis of the results revealed a notable decline in the quantity of detected dead pine trees in Autumn 2023 relative to Spring 2023. In regions where dead trees identified in Spring 2023 had been removed, only a few additional cases were detected in the subsequent autumn monitoring. This indicates that the implemented detection and removal strategy succeeded in curbing the further spread of pine wilt disease, thus demonstrating the practical utility of the proposed monitoring method in disease control efforts.

5. Discussion

5.1. YOLO Model Comparison and Performance Analysis

Through an in-depth analysis of the confusion matrix, this study reveals notable performance differences among the evaluated models in identifying the three target categories: sick_tree, sick_other, and background. For the sick_tree category, the CABiE-YOLOv5 and SEBiE-YOLOv5 models—both based on the YOLOv5 architecture—achieved the highest accuracy, with markedly lower misclassification and omission rates compared to other models. This indicates that the Squeeze-and-Excitation (SE) attention mechanism effectively enhances the extraction of critical visual features associated with infected trees. In contrast, the baseline YOLOv5 model, lacking targeted feature enhancement mechanisms, frequently misclassified sick_tree as background, while YOLOv8 exhibited higher misclassification and omission rates, potentially due to insufficient discrimination of the unique textures and color patterns of infected trees. In the sick_other category, the overall accuracy of all models remained relatively low, with the confusion matrices indicating varying degrees of misclassification and omissions. Nevertheless, SEBiE-YOLOv5 still outperformed other models in this category, delivering higher accuracy and lower error rates. Notably, ECABiE-YOLOv5 performed poorly in distinguishing sick_other from background, resulting in a higher misclassification rate. This further supports the effectiveness of the SE attention mechanism in improving the model’s sensitivity to sick_other and its ability to differentiate it from background. For the background category, most models achieved high accuracy; however, YOLOv8 exhibited a relatively higher misclassification rate, which may stem from overfitting to complex background textures or limited generalization to disease-related features (sick_tree and sick_other).

In summary, the CABiE-YOLOv5, SEBiE-YOLOv5, CBAMBiE-YOLOv5, and ECABiE-YOLOv5 models all incorporate the BiFPN feature fusion network, the EIoU loss function, and different attention mechanisms (CA, SE, CBAM, and ECA) within the YOLOv5 framework. In contrast, the baseline YOLOv5 and YOLOv8 models—despite improvements in the YOLOv8 backbone—do not employ such enhancement strategies. Performance analysis confirms that the SE attention mechanism effectively guides the model to focus on key target regions, thereby significantly improving the detection accuracy of sick_tree and reducing both misclassification and omission rates. In multi-class recognition tasks, baseline models without this mechanism demonstrate comparatively weaker overall performance. Considering detection results across all three categories, SEBiE-YOLOv5 achieves the most balanced and superior performance—not only delivering the best recognition results for the critical sick_tree category, but also maintaining high accuracy for the more challenging sick_other and background categories. This suggests that the model possesses strong generalization ability and adaptability for multi-class forest disease detection tasks.

5.2. Validation and Field Application of SEBiE-YOLOv5

Experimental results from diseased tree detection reveal substantial performance differences among models under varying background complexities. In the simple scenario (Scenario a), all models achieved high detection accuracy, successfully identifying single targets and indicating strong performance in low-interference environments. However, in practical forestry management, interference from objects with spectral characteristics similar to diseased trees is common, making it essential to evaluate model applicability in complex backgrounds (Scenarios b, c, and d). In Scenario b, which included interference from other tree species, models incorporating attention mechanisms accurately identified red-leaf broadleaf trees and effectively excluded interference from spectrally similar objects, whereas the baseline YOLOv5 model misclassified them as infected trees. Notably, SEBiE-YOLOv5 achieved a target confidence score of 90%, significantly higher than the 81% achieved by YOLOv5, indicating that the baseline model exhibits relatively weaker feature extraction capability when dealing with spectrally similar distractors. In Scenario c, which included red soil, SEBiE-YOLOv5 not only correctly identified infected trees but also avoided false detections of healthy species. In contrast, CBAMBiE-YOLOv5 misclassified some healthy trees as sick_other, ECABiE-YOLOv5 misclassified red soil as sick_other, and CABiE-YOLOv5, while detecting the targets, exhibited a higher false detection rate than SEBiE-YOLOv5. In the most complex Scenario d, containing red soil, pine forest, and bamboo forest with three infected trees, SEBiE-YOLOv5 successfully detected all infected trees, whereas all other models missed one. These findings further confirm that SEBiE-YOLOv5 possesses superior overall anti-interference capability, with its attention mechanism effectively enhancing global feature extraction and enabling more precise target recognition in complex background environments.

Based on these findings, SEBiE-YOLOv5 was applied to practical forestry monitoring. In the pine forests of Xishan District, the results indicate that the improved model can more effectively handle interference from similar tree species and spectrally similar features in complex environments. By integrating UAV aerial imagery with the SEBiE-YOLOv5 model, the pixel coordinates of dead infected trees were automatically converted into geographic coordinates, with positional deviations maintained within 10 m. Field verification demonstrated that this UAV-based detection system could accurately extract the locations of dead pine trees, confirming its suitability for large-scale identification and localization in mountainous forest terrains. Based on monitoring data collected in Spring and Autumn 2023, the model achieved verification accuracies ranging from 88.91% to 92.42%. Compared with the multi-scale attention U-Net optimization algorithm proposed by Ye et al. [1], which achieved a recall rate of 57%, the developed model demonstrates a markedly enhanced R performance. Although the precision (P) achieved in this study is slightly lower than the 90% P reported by Wang et al. [31] using an enhanced YOLO model combined with UAV remote sensing, the present model shows a remarkable 89.4% improvement in accuracy under complex scenarios compared with the study by Su et al. [32]. Moreover, the model explicitly addresses common interference factors, such as other dead trees and red-leaf plants, which often affect the identification of PWD-infected pine trees. Field validation during Spring and Autumn 2023 further confirmed that SEBiE-YOLOv5 maintained high detection accuracy while keeping localization errors within an acceptable range. From a practical perspective, the model achieves high accuracy in complex background scenarios while sustaining low positional errors in mountainous forests, thus aligning closely with real-world forestry monitoring requirements. Future research will further explore the model’s performance in detecting PWD-infected dead trees and incorporate additional real-world case studies to provide more efficient technical support for forestry pest and disease management.

5.3. Limitations and Future Work

Despite its demonstrated effectiveness, the current method has two main limitations. First, the discoloration process of PWD—infected pine trees is dynamic. In the early stages of infection, trees typically exhibit no visible color changes, while in the mid-to-late stages, foliage gradually shifts from yellow-green to reddish-brown and eventually to brown [33]. As a result, most existing studies—including this one—primarily focus on detecting trees in the later stages of infection, making early-stage identification particularly challenging. Second, the color characteristics of pine trees are influenced by multiple external factors. For example, drought stress can also lead to foliage discoloration, potentially confounding the accurate detection of PWD [34]. At present, this study is applicable only to potential PWD-affected forest areas, where detection relies on identifying pine trees exhibiting foliage discoloration as an indicator of possible infection. However, it cannot precisely determine the specific cause of such discoloration. To address these challenges, future research should explore two main directions. First, hyperspectral imaging technology could be employed to capture subtle changes in chlorophyll content, thereby improving early-stage detection. Second, multi-temporal data could be incorporated to expand the dataset and support in-depth model optimization, thereby enhancing the model’s ability to differentiate the specific causes of pine discoloration and improving its practical value. The SEBiE-YOLOv5 model proposed in this study is relatively lightweight and therefore well-suited for deployment on mobile or edge devices. This makes it particularly advantageous for meeting the core requirements of PWD prevention and control—namely, early detection and timely intervention—especially in complex forest environments. The method substantially improves monitoring efficiency, reduces labor and material costs compared to manual surveys, and offers a trustworthy technological foundation for swiftly pinpointing and removing dead or infected trees.

6. Conclusions

This study addresses the rapid spread of pine wilt disease and the urgent need for timely identification and removal of infected trees by leveraging UAV-based remote sensing imagery and advanced intelligent recognition techniques. A remote sensing monitoring method tailored for detecting dead pine trees in complex mountainous forest environments was developed. To overcome limitations inherent in traditional monitoring methods—such as missed detections and low recognition accuracy due to small target sizes, complex backgrounds, and interference from visually similar objects—this research integrates multiple feature enhancement and optimization strategies. By embedding advanced attention mechanisms, employing a BiFPN network for efficient multi-scale feature fusion, and introducing the EIoU loss function for improved bounding box regression, the model significantly enhances its ability to perceive key visual features. These enhancements effectively tackle issues like detecting small objects in UAV images, background noise, and accurately pinpointing dead or infected trees in rough terrain. The proposed method offers a scientific and technical foundation for effective pine wilt disease surveillance, enabling rapid, accurate, and cost-efficient monitoring of infected trees. This has substantial significance for forest resource protection and ecological stability. Experimental results demonstrate the following:

(1)

This research utilizes the YOLOv5 object detection model, bolstered by attention modules to sharpen the focus on critical features. It also incorporates BiFPN to facilitate the fusion of multi-scale features and adopts the EIoU loss function to refine the accuracy of bounding box predictions. These improvements boost the model’s precision in identifying dead pines, minimizing both false alarms and overlooked cases. While maintaining high detection precision, the approach significantly boosts the AP and F1. The findings highlight that this approach provides a reliable and budget-friendly option for UAV-assisted remote sensing surveillance, delivering solid technical backing for the accurate and prompt management of pine wilt disease.

(2)

In practical field applications, continuous UAV-based monitoring identified 550 and 66 dead pine trees in the spring and autumn of 2023, respectively. Ground-truth validation confirmed detection accuracies of 88.91% and 92.42%, with spatial deviations between UAV-identified trees and actual ground coordinates ranging from 1 to 10 m. These findings confirm the precision, dependability, and real-world effectiveness of the new monitoring method in rugged mountain forest settings. The remote sensing system crafted in this research offers crucial technical assistance for identifying forest pests and monitoring forest health. By doing so, it plays a vital role in safeguarding regional forest resources and upholding ecological balance.