High-precision individual tree canopy segmentation from UAV remote sensing imagery based on the improved BKFE-UNet model

Canopy segmentation is a crucial step in obtaining canopy parameters in forestry remote sensing, and it holds significant importance for research areas such as forest carbon sequestration. However, canopy segmentation based on UAV imagery still faces challenges including confusion between edges and background, as well as insufficient edge details. To address these issues, this paper proposes a semantic segmentation network based on U-Net which implements boundary and key feature enhancement named BKFE-UNet, which is built upon the U-Net model and integrates a differential boundary attention module (DBM) and a key feature enhancement module (KFEM). The DBM enhances canopy edge information through differential computation, alleviating problems such as adhesion between adjacent canopy edges and inadequate edge details. The KFEM introduces Ghost convolution; by stacking Ghost features, it simultaneously filters out redundant information and enhances the key semantic features of canopy objects, thereby reducing background interference. On the UAV tree canopy segmentation dataset, BKFE-UNet achieves an mPA of 92.62%, an mIoU of 86.50%, an Accuracy of 96.30%, and an F1-score of 0.90, representing improvements of 2.23%, 2.27%, 2.09%, and 0.02, respectively, over the baseline U-Net, demonstrating a significant improvement over the baseline U-Net model.

Accurate canopy segmentation results can provide reliable data support for ecological research such as forest carbon sink estimation, forest health monitoring, and forest resource inventory, and hold significant practical value for maintaining the ecological stability of forest areas and advancing the digital management of forestry ecosystems. Although traditional field surveys can obtain a wealth of on-site canopy parameters, they require substantial time and human resources. Remote sensing technology addresses these issues through its capabilities for large-scale synchronous observation and rapid information acquisition. However, applying remote sensing technology to obtain individual tree canopy parameters must be predicated on individual tree canopy segmentation. Satellite remote sensing suffers from limitations such as low resolution and atmospheric interference, which prevent it from providing sufficient data accuracy for individual tree canopy studies. Unmanned Aerial Vehicle (UAV) remote sensing, a type of low-altitude remote sensing technology, acquires high-resolution imagery with high ground resolution and rich detailed information. Moreover, it is not subject to atmospheric interference during image acquisition, thereby effectively overcoming the shortcomings of satellite remote sensing when applied to individual tree canopy segmentation.

Region-based image segmentation algorithms have been widely applied in early studies of tree canopy segmentation, including the watershed algorithm (Komura et al., 2004; Deng and Li, 2011), region growing algorithm (Theng and Ling, 2008; Ma et al., 2020) and multi-scale algorithm (Jing et al., 2012; Naveed et al., 2019). These methods segment an image by iteratively expanding regions from seed points, enabling effective capture of object boundaries. However, the segmentation process of these algorithms is initiated from treetop seed points, and their performance largely depends on accurate treetop detection. Clustering-based segmentation algorithms, such as k-means (Bunting et al., 2010; Hosingholizade et al., 2025) and Mean-shift (Tusa et al., 2021; Wang et al., 2006), group pixels in an unsupervised manner based on feature similarity, requiring neither prior knowledge nor labeled data. Although clustering methods achieve relatively effective segmentation, they often struggle to cope with irregular canopy shapes and variations in forest density.

With the advancement of deep learning, semantic segmentation methods based on convolutional neural networks have been progressively applied to the task of tree canopy segmentation from UAV imagery. Existing research mainly includes model architectures such as Mask R-CNN, YOLO, Vision Transformer, and U-Net.

Mask R-CNN, building upon Faster R-CNN, achieves semantic segmentation simultaneously with object detection and classification for individual objects (He et al., 2020). The Mask R-CNN algorithm has demonstrated potential for tree canopy identification and segmentation in scenarios such as plantations (Hao et al., 2021), pine forests (Hu et al., 2022), urban woodlands (Ocer et al., 2020), and forest fragments (Braga et al., 2020). YOLO is an object detection algorithm characterized by its compact model size, fast computational speed, and strong generalization capability. YOLO and its upgraded versions have demonstrated application potential in scenarios such as mixed forests (Zhang et al., 2022), orchards (Wei et al., 2024), tropical rainforests (Shi et al., 2025), as well as small and medium-sized trees (Sun et al., 2025). Vision Transformer employs a self-attention mechanism to directly model the relationships between any two image regions, regardless of their spatial distance (Khan et al., 2022). Vision Transformer has demonstrated research value in scenarios such as date palm trees (Al-Ruzouq et al., 2024), horticultural plantations (Vinod et al., 2024), acacia trees (Gibril et al., 2025), and trees outside forests (Jiang et al., 2025).

Mask R-CNN and YOLO are designed for instance-level segmentation, making them prone to topological errors or blurred boundaries when tree canopies are closely adjacent. Vision Transformer, on the other hand, lacks local inductive bias, leading to unstable edge convergence when data is limited. In contrast, U-Net’s symmetric encoder-decoder architecture, coupled with skip connections, effectively preserves high-resolution spatial details and achieves superior segmentation performance even with small-scale annotated UAV datasets. The U-Net model demonstrates excellent performance in the canopy segmentation task (Li R. et al., 2023).

Anagnostis et al. adjusted the input layer of the U-Net architecture to support three-channel and six-channel images and added a dropout layer between the convolutional layers of each block to prevent overfitting on small-scale and homogeneous datasets. Experimental results showed that the adjusted model achieved segmentation accuracies of 0.91, 0.90, and 0.87 on the RGB training, validation, and test sets, respectively (Anagnostis et al., 2021). Li et al. adopted ResNet34 as the backbone network, integrated the CBAM attention module into the skip connections of U-Net, and replaced the cross-entropy loss function with the focal loss function. Experimental results demonstrated that the new model improved mPA by 28.59% and mIoU by 21.82% compared to the baseline model (Li Z. et al., 2023). Chakraborty et al. constructed the UAV-Lichi dataset and selected the U-Net model as the deep learning architecture with ResNet50 as the backbone for tree canopy segmentation research (Chakraborty and Deka, 2023). The model was trained using the Adam optimizer and the Jaccard loss function over 500 epochs. Experimental results showed that the proposed model achieved an mIoU score of 89% and an F1 score of 98% on the UAV-Lichi dataset. During this period, model improvements remained limited to simple backbone replacements and parameter adjustments.

Duan et al. proposed an improved U-Net-based model that employs InceptionV2 to compensate for the limitations of fixed convolutional kernels in capturing multi-scale information. The CBAM module is used to enhance the weights of key features, while the Attention Gate module improves the model’s sensitivity to foreground pixels. Experimental results on the constructed TREE512 dataset showed that overall accuracy (OA), mean precision (mPrecision), mean recall (mRecall), and mean intersection over union (mIoU) reached 90.79%, 88.08%, 89.41%, and 79.9%, respectively, representing improvements of 1.07%, 1.63%, 1.5%, and 2.41% over the baseline U-Net model (Duan et al., 2025). Xu et al. built the ECA-Unet model based on the U-Net architecture, integrating a channel attention mechanism into the decoder to enhance the accuracy and precision of small-scale UAV tree segmentation. The overall accuracy reached 85.87%, with a 1.23% improvement in tree canopy segmentation accuracy compared to the original U-Net model (Xu et al., 2025). Sharifi et al. enhanced the basic U-Net encoder-decoder architecture with residual connections to improve gradient propagation and augmented the attention gate to enable dynamic spatial feature refinement, thereby addressing challenges in urban remote sensing such as tree canopy overlap, classification imbalance, and spectral noise in complex built environments. Experimental results showed that the proposed network achieved an F1 score of 0.91, an IoU of 0.84, a precision of 0.93, and a recall of 0.89, representing improvements of 0.06, 0.10, 0.11, and 0.09, respectively, over the baseline U-Net model (Sharifi, et al., 2025). These works represent significant advances in model improvement methodologies, with careful selection and design of module types and architectural modification positions tailored to specific issues affecting segmentation performance. However, no study to date has specifically targeted the prominent issues of boundary blurring and background confusion in tree canopy segmentation.

Under conditions of dense canopy overlap and complex background interference, the conventional U-Net is prone to issues such as loss of fine details and confusion between foreground and background, indicating room for further optimization. First, U-Net does not incorporate the extraction and utilization of multi-angle composite features. Second, the internal texture and edge structures of tree canopies contain complex semantic information; U-Net can only perform preliminary semantic feature extraction and fails to capture deeper, more critical semantic features. Finally, U-Net cannot effectively leverage prior knowledge that carries strong guiding information, such as object edge cues.

To address these limitations, this research proposes an improved semantic segmentation model based on U-Net. The main improvements made in this study are as follows: (1) Attention modules are added at the skip connections, and feature extraction modules are introduced at the bottom and top of the model, significantly enhancing the U-Net model’s ability to acquire composite feature information, thereby improving segmentation performance. (2) A differential edge attention module is proposed to extract salient edge features from images and incorporate them into training to address the issue of boundary blurring in segmentation. (3) Furthermore, ghost convolution is integrated into the network via the feature extraction module to filter out redundant features while extracting and reinforcing key features, achieving the goal of reducing background confusion in segmentation.

In this study, ablation experiments are conducted to compare segmentation accuracy and model size with and without the addition of the differential edge attention module and the feature extraction module, thereby validating the effectiveness of the improvements and demonstrating the lightweight characteristics of the modules. Through visual presentation of the ablation experiment process, the roles played by the two added modules in addressing the issues of boundary blurring and background confusion are verified. The overall technical roadmap of this study is shown in Figure 1.

Study area

The study area is located within the Millennium Forest in Xiong’an New Area, Hebei Province (39°00′N, 116°02′E), as shown in Figure 2. The study area is located Mountains, characterized by an alluvial plain topography, continental climate featuring semi-humid and semi-arid conditions and distinct four seasons. The study area has an annual mean temperature of 12.4 °C, with historically recorded extreme maximum and minimum temperatures reaching 41.2 °C and −22.2 °C, respectively. The annual sunshine duration is approximately 2,298.4 h, and the annual frost-free period averages about 204 days. Regarding precipitation, the mean annual precipitation is 495.1 mm, though interannual variability is high, with historically recorded extreme maximum and minimum annual precipitation reaching 931.8 mm and 207.3 mm, respectively. The mean annual evaporation is approximately 1,661.1 mm, substantially exceeding the precipitation, reflecting the semi-arid climatic characteristics of the region. There contains over ten species of broadleaf trees and over ten species of coniferous trees in the study area, all of which were even-aged stands in good condition.

Data

Data sources and processing

The data were captured by a DJI M300 RTK drone equipped with a RedEdge-MX camera. The footage was captured on 24 June 2024. On the day of filming, the weather was clear with ample sunlight and virtually no wind. The drone’s flight altitude is 200 m, The spatial resolution of the image is 0.2 m. Each image measures 1,280 × 960 pixels and comprises five bands with red, blue, green, near-infrared, and red edge. Parallel flight lines were used for the mission, with the camera triggered based on distance. Radiometric calibration and dark current correction were performed before collection. A total of 4,000 valid images were acquired, covering an area of 4 km 2. The raw images were saved in TIFF format, and both the flight logs and RTK positioning data at the time of each camera exposure were recorded synchronously.

The images processing is cellebrated in Figure 3. The collected UAV images were processed using Pix4Dmapper software (Pix4D, 2020). Data from the three spectral bands R, G, and B were selected for stitching and synthesized into digital orthophotos. To obtain richer image information, RGB true-color synthesis was performed on the synthesized digital orthophotos. The preprocessed images were uniformly cropped to 128 × 128 pixels. To preserve more complete canopy information, the repetition rate was set to 0.2. An overlap ratio of 0.2 is a commonly used setting in the preprocessing of image tiles (Allen et al., 2024; Maschler et al., 2018).

Datasets

In this paper, a Main Dataset is constructed for model improvement experiments. In addition, an Open Forest Dataset and a Closed Forest Dataset are constructed to validate the generalization performance of the model. Partial data from the three datasets are shown in Figure 4.

Main Dataset: From the segmented canopy images, 138 images were randomly selected to construct the primary dataset. Each image slice was manually annotated using LabelMe software (Wada, 2021) to outline the canopy edges, and the annotation data were exported as binary masks to generate binary label maps distinguishing canopy regions from background areas. Data augmentation was performed by simultaneously applying a 45° clockwise rotation and grayscale transformation to both the data and their corresponding labels, enhancing model robustness in terms of position and color dimensions. The augmented dataset contained a total of 414 high-resolution UAV canopy images, each accompanied by a corresponding binary label map in PNG format. Finally, the data were divided into training, test, and validation sets at a ratio of 8:1:1. The training set was used to fit model parameters, the validation set was used for model selection and hyperparameter tuning during training, and the test set was held back until the model was fully determined and used only for final objective performance evaluation.

Open Forest Dataset: From the segmented canopy images, 174 images of sparse forest were selected. The criterion for identifying a sparse forest image was that approximately more than 70 percent of the image area was covered by tree canopies as judged visually, and the gaps between adjacent tree canopies were sufficiently large to clearly see the underlying ground. Subsequently, the same augmentation and annotation methods used for the primary dataset were applied to process the sparse forest data. The processed sparse forest dataset consisted of a total of 522 high-resolution UAV canopy images and their corresponding binary labels. Finally, the sparse forest dataset was divided into training, test, and validation sets at a ratio of 8:1:1 for model training.

Closed Forest Dataset: From the segmented canopy images, 189 images of dense forest were selected. The criterion for identifying a dense forest image was that approximately more than 70 percent of the image area was covered by tree canopies as judged visually, and the spacing between adjacent tree canopies was so compact that the boundaries of the tree canopies and the forest floor beneath were nearly indistinguishable. Subsequently, the same augmentation and annotation methods used for the primary dataset were applied to process the dense forest data. The processed dense forest dataset consisted of a total of 567 high-resolution UAV canopy images and their corresponding binary labels. Finally, the dense forest dataset was divided into training, test, and validation sets at a ratio of 8:1:1 for model training.

Methods

This chapter aims to address the issues of boundary blurring (Firoze et al., 2023; Wang et al., 2004; Kang et al., 2016) and background confusion (Cheng et al., 2021; Sun et al., 2022; Fan et al., 2024) in tree canopy segmentation of mixed planted forests based on UAV imagery. The ultimate application goal of this method is to provide a high-precision segmentation algorithm for individual tree parameter extraction in forest resource inventory and variable-rate spraying operations in precision forestry. Its practical significance lies in reducing manual survey costs and improving the efficiency of forest resource monitoring through enhanced segmentation quality.

Model structure

Over all structure of BKFE-UNet

U-Net is a commonly used convolutional neural network featuring a symmetrical structure resembling the letter “U”. It is widely used in various scenarios, consisting of encoders, decoders, and jumpers connecting the two. The encoder section is typically constructed by sequentially stacking multiple convolutional modules, and within each convolutional module, features are automatically extracted through the convolutional layer and then activated by an activation function, the activated features enter the pooling layer, progressively abstracting features from the image to extract deeper semantic information. The decoder also consists of multiple convolutional modules that progressively restore the image resolution using upsampling. During the upsampling process, the decoder employs skip connections to fuse feature maps from corresponding resolution levels in the encoder, which integration combines low-level detailed features with high-level semantic information, thereby enhancing the accuracy and refinement of image segmentation.

This research constructs a network architecture based on the U-Net framework named Boundary and Key Feature Enhancement U-Net (BKFE-UNet), as shown in Figure 5. The key components of this network comprise two parts: the differential boundary attention module (DBM), and the key feature enhancement module (KFEM). Firstly, The DBM is embedded in the skip connections between the upsampling and downsampling layers of U-Net, thereby enhancing semantic information with margin-enhanced features. Secondly, KFEM is constructed based on Ghost Convolution, which leverages the ability of Ghost Convolution to eliminate redundant features while retaining key features. Through the multi-layer parallel configuration of Ghost Convolution, the critical information within the canopy is enhanced, thereby reducing the misclassification of background during segmentation. Finally, KFEM is constructed based on Ghost Convolution, which leverages the ability of Ghost Convolution to eliminate redundant features while retaining key features. Through the multi-layer parallel configuration of Ghost Convolution, the critical information within the canopy is enhanced, thereby reducing the misclassification of background during segmentation.

Differential boundary attention module

Edge information serves as valuable prior knowledge, aiding object segmentation by capturing the fine structure and complex shapes of objects (Chen et al., 2026; Wang et al., 2024; Li et al., 2023). However, extracting edge information without performing differentiation may introduce irrelevant information, while considering only edge information may result in the loss of internal features. To mitigate this situation, this research extracts and enhances boundary information prior to performing feature differentiation. This method can simultaneously capture differences in boundary information and obtain differential features through feature differences.

The combination of edge information and feature difference operations effectively delineates the boundaries and contours of segmented objects. Therefore, this study proposes a differential boundary attention module for object segmentation utilizing boundary information.

The module structure is shown in Figure 6. The extraction of edge features is accomplished through three operations: pooling, subtraction, and convolution. This is because edges typically exhibit high gradient values, and applying average pooling helps smooth the features. A first-order difference operator is used to perform differencing operations, where the average-pooled features are subtracted from the original features to highlight edge regions. Following this, the features are passed through a 1 × 1 convolution followed by a Sigmoid activation function to enhance the contrast and saliency of the edge features. The edge features are then re-enhanced through multiplication and addition operations. Finally, the SimAM attention mechanism is utilized to re-perceive important regions, thereby producing the final features. The equations governing these operations are as follows: The equations supporting these operations are shown in Equations 1–3. S i m A M denotes the attention mechanism; A P denotes the 3 × 3 average pooling layer; the C o n v 1 × 1 denotes a convolutional block using the Sigmoid activation function.

The SimAM attention mechanism is a parameter-free attention mechanism. Based on well-established neuroscientific theories, information-rich neurons typically exhibit distinct firing patterns compared to their surrounding neurons and inhibit the latter. An energy function is constructed using the linear separability between a neuron and its neighboring neurons to describe the energy of the current neuron. The greater the linear separability, the lower its energy, and consequently, the more important it is. The energy function for the current neuron is defined as follows: μ ^ and σ ^ 2 represent the mean and variance of all neurons except t. After obtaining the minimum value through Equation 4, the importance of the neuron can be determined using 1 E. Finally, the feature matrix is improved according to Equation 5. E groups all e t * values across both the channel and spatial dimensions.

Key feature enhancement module

To fully extract key features, this research proposes a multi-level key feature extraction mechanism named key feature enhancement module (KFEM). The KFEM stacks the critical information extracted from feature maps at different levels, thereby significantly reinforcing the key features obtained through Ghost convolution while compensating for the information loss incurred during the removal of redundant features.

Ghost convolution (Figure 7) reduces redundant information in the numerous intermediate feature maps of convolutional neural networks, thereby preserving key features while improving model training efficiency.

In practice, given input data X ∈ R c × h × w, where c represents the number of input channels, while h and w denote the height and width of the input data, respectively. The operation of generating n feature maps in an arbitrary convolutional layer can be expressed as: In the formula, * denotes the convolution operation, b is the bias term, Y ∈ R h ′ × w ′ × n is an output feature map with n channels, f ∈ R c × k × k × n is the convolution filter in this layer, Additionally, h ′ and w ′ represent the height and width of the output data, respectively. k × k is the kernel size of the convolution filter.

According to Equation 6, the number of parameters to be optimized (in f and b) is explicitly determined by the dimensions of the input and output feature maps. The output feature maps of convolutional layers typically contain a significant amount of redundancy, with some potentially being similar to each other. Suppose the output feature map is a “ghost” of a few simple transformed intrinsic feature maps. These feature maps are typically small in size and generated by standard convolutional filters. Specifically, the m eigenfeature maps Y ′ ∈ R h ′ × w ′ × m are generated using the main convolutional layer, as shown in Equation 7. f ′ ∈ R c × k × k × n denotes the filter used, where m ≤ n. For simplicity, the bias term has been omitted. The hyperparameters—such as filter size, stride, and padding—are identical to those in standard convolutions (Equation 6) to maintain consistent spatial dimensions (h ′ and w ′) across all feature maps. To further obtain the required n feature maps, a series of simple linear operations is applied to each eigenfeature in Y ′. Generate s ghost features based on the following function. y i ′ is the ith eigenfeature map in Y, In the above function, Φ i, j is the jth linear operation that generates the jth ghost feature map y i j, y i ′ can have one or more ghost feature maps. The final Φ i, s is used to preserve the identity mapping of the intrinsic feature map, as shown in Figure 7b. By utilizing Equation 8, the feature map Y = y 11, y 12, …, y m s of n = m · s can be obtained as the output data of the Ghost convolution.

The structure of KFEM is shown in Figure 8. First, the original feature map undergoes three rounds of downsampling, gradually reducing the spatial resolution to obtain four feature levels at different scales, including the original feature map. This allows the model to simultaneously capture multi-scale contextual information ranging from fine-grained to coarse-grained details. To efficiently extract key features from each level while maintaining computational efficiency, as shown in Equation 9, the model then applies Ghost Convolution to each of the four feature maps, as shown in Equation 10. Subsequently, the key features output by Ghost Convolution from the four levels are first concatenated sequentially along a top-down path, progressively transferring semantic information from the top level to the lower levels with higher spatial resolution, as demostrated in Equation 11. Then, they are concatenated along a bottom-up path, feeding spatial details from the bottom level back to the top level, as demostrated in Equation 12. Through this bidirectional concatenation mechanism, the model ultimately generates two output feature maps rich in key features—one emphasizing high-level semantic information and the other emphasizing high spatial resolution—making them suitable for subsequent vision tasks requiring higher precision. F i denotes the primary feature at the ith level, G i denotes the key feature at the ith level processed by Ghost convolution, C o n c a t e · denotes concatenation of feature maps along the channel dimension, ↓ · denotes downsampling the feature map, ↑ · denotes upsampling of feature maps, Y i denotes output feature of KFEM.

Considering the requirements for model lightweighting, the scale distribution of tree canopy features in UAV imagery, and computational cost, this paper adopts a four-level Ghost convolution structure. A three-level structure, due to its shallower feature hierarchy, suffers from insufficient modeling capacity for the complex texture and boundary information of tree canopies. Although a five-level structure can further extract deep semantic features, it significantly increases the parameter count and computational burden. The four-level structure achieves a better balance among feature representation capability, model complexity, and inference efficiency.

Loss function

Binary cross-entropy (BCE) is one of the widely used loss functions in binary image segmentation tasks, it reflects the direct difference between the predicted mask and the actual label. The definition of BCE can be expressed as Equation 13. In the formula, Y i, j ϵ 0, 1 denotes the true label of pixel, Y ^ i, j ϵ 0, 1 denotes the prediction mask. We employed BCE loss to train the network in this study.

Evaluation indicators

This experiment employs four indicators to evaluate effectiveness of the models: mean Pixel Accuracy(mPA), Accuracy, F1 score and mean Intersection over Union(mIoU).

In the above Equations 14–17, k + 1 is the number of categories, P i i is a Class i pixels predicted as class i. P i j is a pixel of type i, projected to be a pixel of type j.

Statistical analysis was performed using Python 3.9 (SciPy package). The Shapiro-Wilk test was used to assess normality. As the data were not normally distributed, the Wilcoxon signed-rank test was employed for pairwise comparisons of segmentation performance metrics, with a significance threshold of p < 0.05.

Experiment configuration

The model training employed the Adam optimizer with an initial learning rate set to 1 × 10 −4, coupled with a cosine annealing learning rate scheduling strategy to enhance convergence stability. Following the mainstream settings in this field, this experiment adopts the Adam optimizer with an initial learning rate of 1 × 10 −4 and employs a cosine annealing strategy for learning rate scheduling (Alves et al., 2024; Lu et al., 2023). To prevent model overfitting, a Dropout regularization mechanism was introduced into the network, with a dropout probability set to 0.2. To ensure experimental reproducibility, the random seed was fixed at 42. The training batch size was set to 2, and the model was trained for a total of 100 epochs. The experimental hardware environment consisted of an NVIDIA RTX 4060 GPU. During the inference stage, a single-scale conventional inference setting was adopted to balance segmentation accuracy and computational efficiency.

In the ablation experiments and comparative experiments, all participating models were trained under strictly identical parameter settings, including optimizer type, initial learning rate, learning rate scheduling strategy, batch size, number of training epochs, and data preprocessing methods. This ensured that different models were compared fairly under uniform experimental conditions, thereby guaranteeing the authenticity and comparability of the performance evaluation results.

Ablation experiment

To evaluate the effectiveness of each module in this approach, ablation experiments were conducted on the DBM and the KFEM. Firstly, conduct a visual evaluation of the segmentation results. Then, use mPA, Accuracy, mIoU, and F1 score as metrics for assessing model performance.

The results of model accuracy evaluation using mPA, Accuracy, mIoU, and F1 are shown in Table 1, where √ indicates the addition of a module and × indicates no module addition. As shown in the experimental results in Table 1, the individual introduction of DBM and KFEM both led to significant improvements across all accuracy metrics. This is because DBM extracts more canopy edge information from the images, while KFEM enhances key information within the canopy. When both DBM and KFEM are integrated simultaneously, the model accuracy achieves a significant improvement compared to adding only one module. This is because DBM focuses on canopy edge information, while KFEM concentrates on key information within the canopy. The features extracted by the two modules have low overlap and strong complementarity. Therefore, when both modules are added to the model together, the effective information obtained is multiplied, resulting in a substantial enhancement in segmentation accuracy relative to adding either module alone. At this point, the model’s mPA, Accuracy, mIoU, and F1 score improved by 0.15%, 0.19%, 0.28%, and 0.03, respectively. The Wilcoxon signed-rank test revealed that our method achieved significantly higher Dice similarity coefficients than the baseline U-Net (median [IQR]: 0.92 [0.89–0.94] vs. 0.88 [0.84–0.91], p = 0.012). These accuracy metrics indicate that embedding DBM and KFEM into the feature extraction component of the U-Net architecture can effectively enhance model performance.

As shown in Table 2, when DBM is not introduced, the inclusion of KFEM increases FLOPs by only 25.47M (approximately 0.059%) and the number of parameters by only 0.12M (approximately 0.482%). Under the condition where DBM is already introduced, the addition of KFEM similarly brings only a marginal increase of about 25.48M FLOPs and 0.12M parameters, with the increments being highly consistent and both below 0.5%. These results fully validate the design effectiveness of KFEM in filtering redundant information by stacking Ghost features, achieving the dual objectives of feature redundancy reduction and key semantic enhancement at an extremely low resource cost. In summary, the KFEM module provides an efficient feature representation foundation for improving segmentation accuracy while ensuring model lightweighting.

The segmentation results of U-Net and its improved models on medium-sized canopy images are shown in Figure 9. Compared to large-sized canopy imagery, medium-sized canopy imagery exhibits reduced capability in capturing and presenting canopy details, resulting in relatively blurred edge segmentation by the models. The model that incorporates both the DBM and KFEM modules into the U-Net architecture demonstrates improved edge segmentation performance over the baseline U-Net, with the enhancement being primarily evident in the convex portions of the canopy edges, while a portion of these images also exhibit improved segmentation in the concave edge portions. Overall, the performance improvement from model enhancements for medium-sized canopy image segmentation is primarily manifested in the convex regions of canopy edges, and this improvement is relatively diminished compared to that observed in large-sized canopy imagery.

The segmentation results of U-Net and its improved models on tree canopy images prone to background confusion are shown in Figure 10. When using the U-Net model, it is prone to misclassifying grassy areas in the background as small-sized tree canopies. The U-Net models enhanced with either the DBM or KFEM exhibited a significant reduction in misclassification artifacts in their segmentation results. In the BKFE-UNet model integrating both modules, misclassification was completely eliminated.

The segmentation results of U-Net and its improved models on canopy images with closely spaced intervals are shown in Figure 11. The segmentation results of the U-Net model exhibit significant boundary adhesion issues between adjacent tree canopies. The individual incorporation of either the DBM or the KFEM into the U-Net architecture demonstrated no significant improvement in addressing canopy boundary adhesion. The BKFE-UNet model, which integrates both DBM and KFEM modules, effectively reduces boundary adhesion between closely spaced tree canopies, resulting in improved separation of individual tree canopies.

Comparative experiment

As shown in Table 3, this paper compares the performance of multiple traditional image segmentation algorithms, classical semantic/instance segmentation models, and the proposed BKFE-UNet model on the task of UAV-based tree canopy segmentation. The evaluation metrics include mean Pixel Accuracy (mPA), mean Intersection over Union (mIoU), Accuracy, and F1-score.

From the perspective of traditional algorithms, the watershed algorithm achieves relatively the best performance, with an mPA of 75.60%, an mIoU of 37.80%, and an F1-score of 0.73. In contrast, the region growing algorithm exhibits the poorest performance, with all metrics being significantly lower, and an mIoU of only 15.24%, indicating its ineffectiveness in segmenting tree canopies under complex boundary conditions and severe overlap scenarios. Algorithms such as multi-scale segmentation and K-means also yield unsatisfactory results, with mIoU values below 25%. These findings demonstrate that traditional segmentation methods struggle to cope with the complex morphology and dense distribution of tree canopies in UAV imagery. Features extracted by traditional algorithms are identified and summarized based on human experience, which inherently limits the number of features available. Moreover, given the need to manage program design complexity, only one to a few features are utilized in the model, leading to a very sparse amount of information being acquired from the images.

Among deep learning methods, semantic segmentation models such as HRNet, PSPNet, and DeepLab V3+ all achieve superior performance compared to traditional approaches, with HRNet attaining an mIoU of 78.35% and an F1-score of 0.80. The instance segmentation model Mask R-CNN and the single-stage model YOLOv5n-seg further improve segmentation accuracy, achieving mIoU values of 80.15% and 80.64%, and F1-scores of 0.85 and 0.86, respectively. These results demonstrate the significant advantages of deep learning models in the task of tree canopy segmentation. This is attributed to deep learning’s superior nonlinear mapping capacity and feature self-learning capability, which allow it to effectively mine the abundant contextual information embedded in images.

As a representative model of the encoder-decoder architecture, U-Net achieves excellent performance, with an mPA of 90.39%, an mIoU of 85.23%, and an F1-score of 0.88, significantly outperforming the aforementioned comparison models. This is because U-Net’s encoder-decoder architecture, combined with skip connections, can simultaneously preserve high-resolution spatial details and deep semantic features, while requiring no positional assumptions about the targets and being free from the constraints of candidate regions or grid partitioning. The proposed BKFE-UNet achieves the best results across all evaluation metrics: an mPA of 92.62%, an mIoU of 86.50%, an Accuracy of 96.30%, and an F1-score of 0.90, representing improvements of 2.23%, 2.27%, 2.09%, and 0.02, respectively, over the baseline U-Net. These results fully validate the effectiveness of the introduced DBM and KFEM in enhancing edge information, filtering redundant features, and improving segmentation accuracy.

Generalization performance evaluation

Two datasets, Open Forest Dataset and Closed Forest Dataset, were constructed using dense forest imagery and sparse forest imagery, respectively, to evaluate the generalization performance of the model. The tree canopy segmentation accuracy of the model on the open forest dataset and closed forest dataset is shown in Tables 4, 5, respectively. It can be observed that the BKFE-UNet model proposed in this study achieves superior performance under both dense and sparse forest stand densities, demonstrating that the proposed model possesses strong generalization capability.

To address the common issues of blurred segmentation edges and background confusion in UAV based tree canopy segmentation tasks, this study proposes an improved U-Net based model named BKFE-UNet.

This method aims to extract more boundary information of tree canopies while enhancing key features of the segmented objects and reducing redundant information, thereby achieving clearer edges and fewer background misclassifications. To this end, two modules were introduced based on the U-Net model: the Boundary Depth Information Capture Module (DBM) for capturing detailed boundary information and the Key Feature Enhancement Module (KFEM) for enhancing key features. When the DBM was introduced alone, the model improved its mPA, mIoU, Accuracy, and F1 score by 1.19%, 1.00%, 0.95%, and 0.01, respectively. When the KFEM was introduced alone, the model improved its mPA, mIoU, Accuracy by 0.66%, 0.26%, 0.51%, respectively. A comparison shows that the DBM contributed more to model accuracy improvement than the KFEM. The reason for this difference may be that the DBM plays a more direct role in extracting tree canopy boundaries, whereas the Ghost convolution used in the KFEM fails to preserve sufficient effective information when removing redundant features, thus limiting the efficiency of key feature enhancement. The specific functions of the two modules within the model require further investigation.

When only the DBM was introduced, the model’s FLOPs increased by 1,131.21 M and the Params increased by 0.86 M. When only the KFEM was introduced, the model’s FLOPs increased by 25.47 M and the Params increased by 0.12 M. The increase in model computational complexity and the number of parameters brought by the KFEM was significantly lower than that of the DBM, highlighting the lightweight nature of the KFEM. This is because, considering the complexity of real world application scenarios and hardware limitations, a lightweight convolutional architecture, Ghost Convolution, was chosen when building the key feature enhancement module. After stacking a limited number of layers, the computational complexity and model size did not increase significantly, while the segmentation accuracy was effectively improved.

The proposed model BKFE-UNet achieved improved accuracy on the validation set compared to the U-Net model. The model’s mPA, Accuracy, mIoU, and F1 score increased by 2.23%, 2.27%, 2.09%, and 0.02, respectively. From the visualization of segmentation results, it can be observed that for medium sized tree canopies with complex edges, the BKFE-UNet model significantly optimized edge segmentation compared to the baseline U-Net model. For canopy images with complex backgrounds, the BKFE-UNet model noticeably reduced background misclassification. For images with closely spaced tree canopies, the BKFE-UNet model markedly improved the edge adhesion between adjacent tree canopies. Overall, the model effectively addressed the optimization of tree canopy edge segmentation and the issue of background confusion.

In addition, the model achieved an mPA of 93.58%, an mIoU of 88.81%, an Accuracy of 96.42%, and an F1 score of 0.92 on the Open Forest Dataset. On the Closed Forest Dataset, the model achieved an mPA of 90.21%, an mIoU of 81.73%, an Accuracy of 93.75%, and an F1 score of 0.88. The segmentation accuracy of the model on the Open Forest Dataset was higher than that on the Main Dataset. This is because the tree canopies in the open forest are farther apart from each other, and the shape of individual tree canopies is relatively regular, leading to higher segmentation accuracy. The segmentation accuracy of the model on the Closed Forest Dataset was lower than that on the Main Dataset. This is because the tree canopies in the closed forest are more densely packed, making it easier for misclassification to occur between adjacent tree canopies. Compared to the baseline U-Net, BKFE-UNet improved mPA by 1.35%, mIoU by 1.30%, Accuracy by 1.28%, and F1 score by 0.02 on the Open Forest Dataset. On the Closed Forest Dataset, compared to the baseline U Net, BKFE-UNet improved mPA by 2.08%, mIoU by 0.86%, Accuracy by 1.18%, and F1 score by 0.02. These results indicate that the proposed BKFE-UNet model can achieve improved segmentation accuracy on datasets with different tree canopy densities, demonstrating the model’s strong generalization performance.

The BKFE-UNet model proposed in this study incorporates the DBM and KFEM into the base U Net model, improving the overall segmentation accuracy. Among the two modules, the DBM contributes more to the improvement of segmentation accuracy, while the KFEM exhibits lightweight characteristics. Visualization results of tree canopy segmentation show that the BKFE-UNet model effectively addresses the edge and background issues in tree canopy segmentation. Accuracy results on the Open Forest Dataset and Closed Forest Dataset demonstrate that BKFE-UNet achieves strong generalization performance across different tree canopy scenarios. However, the internal mechanism by which the model optimizes edge and background problems remains unclear and requires further investigation.

This research presents an enhanced U-Net model called BKFE-UNet integrated with the DBM and the KFEM for canopy segmentation of UAV imagery. The DBM extracts semantic features of canopy edges from the images, while the KFEM integrates the key features extracted through Ghost Convolution, thereby more effectively leveraging its advantage of removing redundant features while preserving critical information. The proposed BKFE-UNet achieved remarkable performance on the UAV canopy segmentation dataset with mPA, Accuracy, mIoU, and F1-score reaching 92.62%, 96.30%, 86.5%, and 0.92 respectively, demonstrating consistent improvement over the baseline U-Net model. In terms of visual segmentation results, the BKFE-UNet model shows significant enhancements in edge details and reduced boundary adhesion between adjacent objects. The improvements made in this study, particularly the lightweight nature of KFEM, were validated through ablation experiments on model size. The generalization capability of the model under different scenarios was verified through experiments on two datasets: the Open Forest Dataset and the Closed Forest Dataset. The findings of this study can provide a high-quality segmentation algorithm for forest resource inventory and precision forestry, thereby improving the efficiency of forest resource monitoring. However, the specific internal mechanisms by which the model optimizes the two issues of edge blurring and background confusion remain unclear and require further investigation.