Identification of Cypress Bark Beetle-Infested Cypress Based on LiDAR and RGB Imagery

Abstract

Forest pests and diseases are some of the major disturbances affecting the stability of forest ecosystems. Accurate identification of insect-infested trees is therefore crucial for assessing forest health and implementing precision forestry management. This study focuses on stand-level detection of cypress trees (Cupressus funebris Endl.) that were affected by the cypress bark beetle (Phloeosinus aubei Perris), and the framework enables individual tree segmentation, insect-infested tree detection, and stand infestation assessment. Firstly, individual trees were extracted from Light Detection and Ranging (LiDAR) point cloud data using the layer-stacking seed point algorithm. Based on the segmented tree crowns, four vegetation indices (Visible Atmospherically Resistant Index (VARI), Visible-band Difference Vegetation Index (VDVI), Red-Green Index (RGI), and Color Index of Vegetation Extraction (CIVE)) were calculated from Unmanned Aerial Vehicle (UAV) RGB imagery. Insect-infested cypress trees were extracted through threshold segmentation. Through visual interpretation, the optimal vegetation index was determined and the infestation rate at the stand level was calculated. Based on the above framework, a total of 1368 trees were identified in the cypress stand, with a segmentation Precision of 82.51%, a Recall of 80.00%, and an F1-score of 81.24%. RGI achieved the best performance (Precision = 100.00%, Recall = 86.96%, F1-score = 93.02%) and identified 20 infested trees, accounting for 1.46% of the cypress stand. Supplementary experiments further confirm the superiority of the RGI index and the μ ± 2σ thresholding method. These results demonstrate that the proposed method enables rapid detection of the infested cypress trees, effective monitoring of stand health and infestation severity, thereby supporting informed decision-making in pest control and forest management.

1. Introduction

Forests are a key component of the ecosystems on Earth; they play a crucial role in preserving biodiversity, regulating climatic conditions, and conserving water and soil resources [1,2,3]. The health status of forest ecosystems directly impacts the equilibrium of ecological processes and affects the stability of the overall ecological framework [4,5]. Plant diseases and insect pest attacks are among the most prevalent natural disturbances in forest ecosystems, leading to landscape-level tree mortality, and escalating into severe threats to the stability and biodiversity of the forest ecosystem [6,7,8]. Under these challenging environmental conditions, the systematic monitoring and detection of plant diseases and insect pests have become critical priorities within forest health management frameworks [9,10]. Specifically, the precise identification of insect-infested trees constitutes a foundational prerequisite for conducting fine-scale forest health assessments and enabling the implementation of targeted pest management strategies [8].

Traditional approaches for monitoring insect-infested trees are often inefficient, costly, and time consuming [11]. The development of remote sensing technology offers an efficient and practical alternative [12]. The most common approaches include threshold segmentation using vegetation indices (VIs) [13,14,15] and machine learning (ML)-derived methods [16,17,18]. Traditional ML methods, such as support vector machine (SVM) and Random Forest (RF) [19,20,21], primarily capture insect-infested signals at a pixel scale, therefore are commonly used for extracting infested regions. Deep-learning models, such as Faster R-CNN and YOLO [21,22], are capable of detecting and delineating individual insect-infested trees, but generally require large amounts of labeled training data and substantial computational resources [23]. While transfer learning or one-class anomaly detection can mitigate sample scarcity, the performance is sensitive to the specific distribution of training samples and local domain characteristics. On the contrary, index-based thresholding strategies rely directly on spectral response characteristics and provide a more transparent and physically interpretable framework, which is advantageous for stand-scale assessment and practical application [24].

There are two procedures in threshold segmentation by VIs, namely individual tree segmentation and VIs extraction, and an accurate individual tree delineation is a prerequisite for precise insect-infested tree identification. The common data source for individual tree segmentation is Unmanned Aerial Vehicle (UAV)-based RGB images and LiDAR data [25,26,27]. As RGB imagery mainly captures horizontal spatial information, its performance in individual tree detection is strongly affected by illumination variation, canopy shadowing, crown occlusion, and understory complexity in structurally complex forests, leading to unstable segmentation results and detection accuracy [28]. Compared with RGB imagery, LiDAR data provide high-ranging resolution, strong interference resistance, and canopy penetration signals through multiple-return measurements, which allows more accurate three-dimensional characterization of forest structure and individual tree delineation [29]. Current LiDAR-based individual tree segmentation methods mainly include Canopy Height Model (CHM) seed point algorithms [30], Point Cloud Segmentation (PCS) [31], and Layer Stacking Segmentation (LSS) seed point algorithms, and so on [32]. Using three individual tree detection methods based on LiDAR data (PointNet++, PCS, and LSS), Liu et al. [33] found that all three methods achieved accuracies above 86% and were effective in reliably identifying individual trees. Using four segmentation methods (CHM, pit-free CHM, PCS, and LSS), Yang [34] compared their performances across forests of varying types and densities, and the results show that the LSS method demonstrated the highest stability regardless of forest type or density.

While LiDAR data provide accurate individual tree delineation results, the detection of insect-infested trees largely depends on spectral differences in tree canopies, which can be effectively characterized using VIs. Although multispectral VIs methods, e.g., Normalized Difference Vegetation Index (NDVI) [35,36] and Enhanced Vegetation Index (EVI) [37], have been widely applied and provide rich spectral information, their application on UAV platforms is limited by sensor cost and accessibility, whereas RGB imagery is more widely available and commonly used in forests’ daily monitoring [38]. Consequently, exploring RGB-based methods have greater potential for practical applications. With Red-Green and Blue-Green band ratios, Yu et al. [39] developed an iterative statistical analysis method to automatically detect individual insect-infested pine trees, and achieved a satisfactory result that both Recall and Precision reached approximately 95%. Using three RGB indices (CIELAB color space (a*, b*), Green Area (GA), Normalized Green-Red Difference Index (NGRDI), and Triangular Greenness Index (TGI)), Sancho-Adamson et al. [40] successfully distinguished Verticillium-infected olive plants from healthy controls during the asymptomatic stage. With the Extra-Green index, Li et al. [41] categorized the study area according to the degree of infection severity from pine wood nematode disease. Overall, RGB VIs are effective in distinguishing insect-affected trees from healthy ones.

To develop an efficient and practical technical approach for insect-infested cypress identification, this study used LiDAR point cloud data and UAV RGB imagery for individual tree segmentation, infested-tree identification, and infestation rate estimation at the forest stand level. The proposed workflow aims to support rapid identification of infested trees through a one-time tree inventory combined with periodic monitoring cycles, trying to provide a practical foundation for timely forest stand management and decision-making in operational contexts.

2. Study Area and Data

2.1. Study Area

The study area is located in Chengdu Longquan Mountain Urban Forest Park, Sichuan, China (Figure 1). The park covers an area of 1275 km² and is located between 104°05′38″–104°36′17″ E and 30°12′29″–30°57′14″ N. The park is in a subtropical humid climate region that receives an average annual precipitation of 895.6 mm and about 83% of it occurs during May to September. The annual sunshine of the park is around 1036.8 h and the frost-free period lasts for about 353 days in a year. The ecosystem exhibits a relatively fragile structure and function, and the vegetation type is monotonous, dominated by secondary plant communities such as Chinese cypress (Cupressus funebris Endl.), eucalyptus (Eucalyptus spp.), and oak (Quercus spp.), with cypress accounting for over 80% [42].

According to the field survey conducted in May 2024, we found that the cypress is infested by cypress bark beetles (Phloeosinus aubei Perris) and the infested cypress trees were scattered and distributed in the northern part of the park. We chose a representative cypress bark beetle-infested cypress stand as our study area with reference to the “One Map of Forest Land” data for Longquan Urban Forest Park. The Digital Orthophoto Map (DOM) of the study area was acquired in July 2024 and the details are shown in Figure 1a. Based on crown color and morphological characteristics, the cypress was classified as being in the third stage of red-attack development. To evaluate the propagation and reliability of the proposed method, another affected cypress stand was chosen one year later; details of the DOM (acquired in September 2025) are shown in Figure 1b. Based on crown color and morphological characteristics, the cypress was classified as being in the fourth stage of gray-attack development. It should be noted that no management activities such as afforestation, tending, or logging were conducted within the forest stand during 2021–2025. The investigated stands consist of mature to over-mature cypress forest stands, where structural changes are generally gradual.

2.2. Data Acquisition

2.2.1. Aircraft LiDAR Data

LiDAR data of the entire Chengdu Longquan Mountain Urban Forest Park were collected in July 2021 using the RIEGL LMS-Q680i LiDAR sensor (RIEGL Laser Measurement Systems GmbH, Horn, Austria) that was carried by Kodiak 100 aircraft (Daher, Sandpoint, ID, USA). Four flights were conducted to cover the whole study area, and the aircraft flew at an average altitude of approximately 2000 m. The LMS-Q680i is a new-generation laser scanner that supports a maximum laser pulse repetition rate of 400 kHz, a 60° scan angle, a scanning rate of 200 lines per second, and a ranging accuracy of 20 mm. In this study, the airborne LiDAR data acquired in 2021 were used for individual tree delineation and crown boundary extraction.

The airborne LiDAR point cloud data preprocessing steps included point cloud mosaicking, noise removal, ground-point classification, and height normalization using ground points. The dataset was reprojected to the WGS_1984_UTM_Zone_48N coordinate system to ensure consistency with the RGB data.

2.2.2. UAV RGB Imagery

Based on the “One Map of Forest Land” dataset of Longquan Urban Forest Park, a representative cypress bark beetle-infested cypress stand was selected as the study area. The RGB imagery was acquired in July 2024 using a DJI Phantom 4 RTK drone with an RGB camera (SZ DJI Technology Co., Ltd., Shenzhen, China). The UAV flew at an altitude of 100 m and has an effective image resolution of 4864 × 3648 pixels. The onboard camera features a 1-inch CMOS sensor with a focal length of 8.8 mm, a field of view of 84°, and a mechanical shutter, providing high-quality imagery with reduced motion distortion. The ground sampling distance was approximately 4.48 cm, enabling fine-scale representation of individual tree crowns.

The collected images were imported into DJI Terra (v4.3.0) to verify data completeness, after which image mosaicking was performed to generate a DOM using the visible-light processing module. The projection coordinate system was set to WGS_1984_UTM_Zone_48N. The resulting DOM was then spatially registered with the LiDAR data. The co-registration accuracy was assessed using 15 control points, yielding a horizontal RMSE of approximately 0.16 m, which is less than 5% of the average crown diameter. The registered DOM is displayed in Figure 2, and the area within the red line () represents the selected cypress forest stand.

3. Methods

In this study, trees were delineated from LiDAR point cloud data, four VIs were calculated at the individual tree level from UAV RGB imagery, and a mean ± two standard deviations threshold was applied for the identification of the infested trees. The infestation rate on the stand level was then calculated. Details are listed in the following sections.

3.1. Individual Tree Segmentation Algorithm

Individual tree segmentation was performed using the LSS method in LiDAR360(v5.2). The LSS approach consists of two main stages. In the first stage, the LiDAR point cloud was vertically sliced into multiple height layers, ranging from 1 m above the ground to the maximum tree height. Within each layer, an adaptive-window local maximum filter was applied to identify potential tree tops, which were defined as seed points representing individual tree locations. These seed points were iteratively refined using K-means clustering based on the distance between points and seed locations until the seed positions converged. In the second stage, the point cloud was segmented based on the finalized seed points using the PCS algorithm. For each height layer, Thiessen polygons were constructed from the seed points to partition the point cloud. The polygons from all layers were subsequently merged to generate the final individual tree segmentation results, producing representative tree crown outlines.

After segmentation, the location and crown width of each tree were obtained. Finally, the spatial extent of each tree was generated using the center coordinates and crown width.

3.2. Evaluation Indices for Individual Tree Segmentation

Validation of individual tree segmentation is typically performed using visual interpretation from RGB data. Segmentation accuracy was assessed using the Intersection of Union (IoU) metric. An IoU value of ≥0.4 was regarded as True Positive (TP), while values < 0.4 were considered False Positive (FP); False Negative (FN) refers to trees that existed in reality but were not detected in the segmentation results. Based on these results, Precision, Recall, and F1-score [43,44] were calculated to evaluate the performance of the point-cloud-based individual tree segmentation. Precision measures the proportion of correctly identified trees among all detected trees. Recall indicates the proportion of actual trees that were successfully detected. The F1-score, ranging from 0 to 1, is the harmonic mean of Precision and Recall and provides an integrated measure of segmentation performance. Higher values indicate better accuracy. The formulas are as follows:

$$\mathit{Precision}={\displaystyle \frac{TP}{TP+FP}}\times 100\%$$

(1)

$$\mathit{Recall}={\displaystyle \frac{TP}{TP+FN}}\times 100\%$$

(2)

$$F1-\mathrm{score}=2\times {\displaystyle \frac{\mathit{Precision}\times \mathit{Recall}}{\mathit{Precision}+\mathit{Recall}}}$$

(3)

3.3. Identification of the Infested Tree

Four VIs based on RGB data were used to test the performance in infested tree identification, namely Visible Atmospherically Resistant Index (VARI), Visible-band Difference Vegetation Index (VDVI), Red-Green Index (RGI), and Color Index of Vegetation Extraction (CIVE). The calculation formulas are described below.

$$\mathit{VARI}={\displaystyle \frac{G-R}{G+R-B}}$$

(4)

$$\mathit{VDVI}={\displaystyle \frac{2G-R-B}{2G+R+B}}$$

(5)

$$\mathit{RGI}={\displaystyle \frac{R}{G}}$$

(6)

$$\mathit{CIVE}=0.441\times R-0.811\times G+0.385\times B+18.78745$$

(7)

where R, G, and B denote the reflectance of the red, green, and blue bands, respectively.

Larger VARI and VDVI values usually imply greener vegetation and a healthier status; smaller or even negative values imply an unnormal status or even insect-infested trees. On the contrary, for RGI and CIVE, a smaller value usually implies a healthier vegetation status, while larger values indicate an unhealthy status.

After calculation of the four VIs at the pixel scale, they were averaged at the tree crown scale, and the threshold for identifying insect-infested trees was determined using the mean ± two standard deviations (μ ± 2σ), which is widely applied in biological and environmental studies. This approach is simple and robust and reflects both the central tendency and variability of the data. Trees with VI values outside the threshold range were classified as infested. Specifically, for the VARI and VDVI indices, trees with values lower than μ − 2σ were identified as insect-infested trees. In contrast, for the RGI and CIVE indices, trees with values higher than μ + 2σ were classified as insect-infested trees.

The results were then compared with visual interpretation data from RGB imagery. Based on this comparison, Precision (see Formula (1)), Recall (see Formula (2)), and F1-score (see Formula (3)) were calculated to evaluate the performance and applicability of the different VIs for insect-infested tree detection.

3.4. Estimation of the Infestation Rate on Stand Level

After completing individual tree segmentation and insect-infested tree identification, the spatial localization of insect-infested trees and the quantitative calculation of the infestation rate were further conducted. On this basis, the number of insect-infested trees (N_d) from the threshold discrimination results of the optimal VIs and the total number of individual trees (N_t) from LiDAR point cloud segmentation were calculated, and the infestation rate was derived according to the following formula:

$$\mathrm{Infestation} \mathrm{Rate}={\displaystyle \frac{N_d}{N_t}}\times 100\%$$

(8)

4. Results

4.1. Results and Accuracy Assessment of Individual Tree Segmentation

A total of 1368 trees were identified using LiDAR point cloud segmentation; the vertical and spatial distribution is displayed in Figure 3 and Figure 4, respectively. Each circle represents the crown area. To evaluate the accuracy of the segmentation results, two sample plots (30 m × 30 m) were randomly selected for visual interpretation (Figure 4).

Based on the results from visual interpretation of the sample plots (

Table 1. Accuracy for visual interpretation of individual tree segmentation based on LiDAR data.

	TP	FP	FN	Precision	Recall	F1-Score
Sample plot 1	85	21	15	80.19%	85.00%	82.52%
Sample plot 2	99	18	31	84.62%	76.15%	80.16%
Total	184	39	46	82.51%	80.00%	81.24%

), a total of 184 trees were correctly identified, 39 were misclassified, and 46 trees were omitted. Accordingly, the Precision of individual tree segmentation was 82.51%, the Recall was 80.00%, and the F1-score was 81.24%. The commission and omission mainly occur in areas where trees are tall and densely distributed. In these areas, interconnected tree crowns formed a continuous canopy, which adds difficulty to the segmentation. In addition, the aircraft operates at an altitude of approximately 2000 m, resulting in a low point cloud density. Nevertheless, the segmentation performance is satisfactory and provides a reliable basis for the subsequent infested-tree identification.

4.2. Extraction of Infested Trees from Different VIs

Four VIs (VARI, VDVI, RGI, and CIVE) were calculated at the crown area of each tree and the results are displayed in Figure 5. For VARI and VDVI, smaller values represent a healthier status, while larger values indicate less greenness, which is a sign of potential infection. For RGI and CIVE, in contrast, larger values imply a healthier state, while smaller values represent an unhealthy condition. Despite differences in their calculation formulas, the four vegetation indices consistently highlighted unhealthy cypress at similar spatial locations. Trees with larger crown areas generally exhibited higher greenness values, whereas those with yellow or orange tones were typically associated with smaller crowns. This pattern likely reflects differences in tree vigor, as individuals with poor crown condition may have lower resistance to biotic stress and are therefore more susceptible to cypress bark beetle infestation.

4.3. Threshold Segmentation for Infested Trees

The scatterplots of the VIs of each tree and the threshold segmentation were displayed in Figure 6. Healthy trees are represented with a blue color and the infested trees with red.

The average VARI value is 0.32 and ranges from −0.20 to 0.40. VARI values smaller than 0.22 are considered infested and there is a total of 21 trees identified.

The average VDVI is 0.24 and ranges from −0.05 to 0.30, and values less than 0.18 are considered infested. There is a total of 24 trees identified. As can be seen, the value distribution of VDVI is similar to that of VARI, but the infested results are slightly different in number and location.

The RGI values range from 0.58 to 1.26, with an average of 0.66, primarily concentrated between 0.58 and 0.76. Outliers occurred at higher values (>0.76); there are 20 trees identified. The distribution of healthy trees is clustered and the location of the infested trees is similar to that of VARI.

The CIVE values exhibited the widest range from −34.10 to 28.20, with a mean value of −16.40. Those with CIVE values ≥ −3.85 were identified as infested and there are 26 trees identified. The CIVE distribution is scattered and the method recognized the largest number of infested trees.

The identification results were examined by visual interpretation from the RGB image; the spatial distributions are presented in Figure 7 and the evaluation results are summarized in

Table 2. Accuracy assessment comparison of cypress bark beetle-infested tree identification based on different VIs.

VIs	TP	FP	FN	Precision	Recall	F1-Score
VARI	20	1	3	95.24%	86.96%	90.91%
VDVI	20	4	3	83.33%	86.96%	85.11%
RGI	20	0	3	100.00%	86.96%	93.02%
CIVE	20	6	3	76.92%	86.96%	81.63%

.

There are 23 infested trees from visual interpretation and 20 of them are identified by all four indices (TP). The omission errors (FN) for all four indices were primarily due to the omission in the individual tree segmentation stage, which is an error propagation. The largest divergence is the commission errors (FP), where the RGI index has the best performance and the other three indices are all overestimated. VARI, VDVI, and CIVE resulted in 1, 4, and 6 commission errors, respectively, most of which occurred near the canopy shadow. Among them, CIVE exhibited a more pronounced sensitivity to shadow, resulting in the highest commission rate. In addition to boundary effects, VDVI and CIVE were further affected by mixed-species trees and background vegetation, which contributed to their increased commission errors.

In summary, the RGI index achieved the highest Precision of 100%, effectively avoiding commission errors, and achieved an F1-score of 93.02%. The VARI index demonstrated balanced performance, with a Precision of 95.24% and an F1-score of 90.91%, indicating robust overall recognition accuracy. The VDVI index exhibited lower Precision and F1-score than VARI, with a Precision of 83.33% and an F1-score of 85.11%, reflecting greater susceptibility to interference from complex canopies and non-target vegetation. The CIVE index had the highest FP rate and the lowest overall accuracy, with a Precision of 76.92% and an F1-score of 81.63%, making it the least suitable for infested tree identification. The Recall for all four VIs was 86.96%, primarily constrained by the accuracy of individual tree segmentation.

4.4. Estimation of Infestation Rate

Following individual tree segmentation based on LiDAR point cloud data, a total of 1368 trees (N_t) were identified in the selected cypress forest stand. Subsequent infested tree identification using RGB imagery, based on the optimal VIs (RGI) with the highest F1-score, 20 infested trees were identified (N_d). The forest stand infestation rate can be calculated based on Formula (8) and reaches a result of 1.46%.

This result implies that the stand is generally in a healthy state, with a low proportion of infested trees, and has not yet reached the threshold for disaster classification, i.e., 10%. However, the cypress bark beetle has strong reproductive capacity and dispersal potential. Under favorable environmental conditions, its population size and the associated damage can increase rapidly. Therefore, continuous and regular monitoring remains necessary even though no severe damage has been observed at this stage. Close observation of infestation dynamics is required to prevent a shift from sporadic occurrences to large-scale outbreaks.

5. Discussion

5.1. The Reliability of the RGI Index

To test the propagation of the RGI index and its reliability at different infestation stages, the four VIs were compared again in another experiment conducted one year later under another forest stand, a different infestation stage, and different sensors. Another infested cypress stand of Longquan Mountain Urban Forest Park was selected. The RGB data were acquired in September 2025, with a DJI M300 RTK drone equipped with a Zenmuse P1 sensor (SZ DJI Technology Co., Ltd., Shenzhen, China). The detail of the DOM is shown in Figure 1b; the color of infested trees was gray, which is the final stage of the infestation, and the color was apparently different from the previous stage, which is displayed in Figure 1a. Similar data processing was conducted and the applicability of the VIs is shown below.

There are a total of 768 trees identified and the scatter plots of the four VIs and segmentations are shown in Figure 8. The spatial distribution of the evaluation result is shown in Figure 9 and the statistical information is listed in

Table 3. Accuracy comparison of cypress bark beetle-infested tree identification results based on μ ± 2σ threshold segmentation for different VIs.

VIs	TP	FP	FN	Precision	Recall	F1-Score
VARI	38	1	10	97.44%	79.17%	87.36%
VDVI	22	3	27	88.00%	44.90%	59.46%
RGI	38	0	10	100.00%	79.17%	88.37%
CIVE	38	2	10	95.00%	79.17%	86.36%

. Despite different segmentation thresholds, the comparison results of the two experiments are largely consistent. RGI also achieved the best performance among the four indices, with an accuracy of 100.00% and an F1-score of 88.37%. VARI also ranked second, with an accuracy of 97.44% and an F1-score of 87.36%. The CIVE index showed improved performance compared with the previous experiment, achieving an accuracy of 95.00% and an F1-score of 86.36%. The Recall rate for RGI, VARI, and CIVE was the same, which is 79.17%. On the contrary, VDVI exhibited a significant decline in performance. Its accuracy is 88.00%, the Recall rate is only 44.90%, and the F1-score is 59.46%. This omission is mainly attributable to illumination and shadow effects, together with the grayish appearance of the infested cypress, which weakened the sensitivity of VDVI. Overall, these comparisons demonstrate that RGI exhibited the greatest stability across different forest stand, infestation stages, and sensor platforms.

5.2. The Reliability of the Thresholding

To further examine the stability of the μ ± 2σ threshold-based strategy, the InterQuartile Range (IQR) method was introduced for comparison. The IQR criterion defines potential outliers using Q3 + 1.5 × IQR or Q1 − 1.5 × IQR and is generally less influenced by extreme values than moment-based statistics, where Q1 and Q3 represent the 25th and 75th percentiles of the data, and IQR = Q3 − Q1. The IQR thresholding strategy was also applied to the VIs of the cypress stand illustrated in Figure 4. For a more straightforward comparison with the μ ± 2σ thresholding strategy, the result from IQR-based classification is also presented in scatter distribution (Figure 10), and the quantitative results are summarized in

Table 4. Accuracy comparison of cypress bark beetle-infested tree identification results based on IQR threshold segmentation for different VIs.

VIs	TP	FP	FN	Precision	Recall	F1-Score
VARI	20	4	3	83.33%	86.96%	85.11%
VDVI	20	8	3	71.43%	86.96%	78.43%
RGI	20	5	3	80.00%	86.96%	83.33%
CIVE	20	5	3	80.00%	86.96%	83.33%

.

The classification results derived from the μ ± 2σ strategy and IQR strategy are shown in Figure 6 and Table 2, Figure 10 and Table 4, respectively. Both approaches produced generally consistent Recall values, and differences were observed in Precision and F1-score. Under the μ ± 2σ rule, RGI achieved the highest F1-score (93.02%) with no false positives, whereas the IQR-based method resulted in lower Precision and F1-scores across all indices. For example, the F1-score of RGI decreased from 93.02% to 83.33% under the IQR criterion. Similar trends were observed for VARI and VDVI. These results indicate that the μ ± 2σ threshold provides a more balanced trade-off between false positives and false negatives. The IQR-based method, although less influenced by extreme values, tend to generate more conservative or variable detection outcomes in our case.

5.3. Uncertainties and Prospects

Despite the satisfactory and stable performance of RGI across different study areas, infestation stages, and sensors, the overall identification accuracy was primarily constrained by the individual tree segmentation derived from LiDAR data. The LiDAR data used in this study were acquired from an aircraft at an altitude of approximately 2000 m to cover a large area (over 1000 km²), which resulted in a relatively low point cloud density and, consequently, limited the accuracy of individual tree segmentation. We believe that with a higher density of LiDAR point cloud data by UAV, better segmentation accuracy can be achieved [45]. In addition, there is a temporal inconsistency between the LiDAR data (acquired in 2021) and the RGB imagery (acquired in 2024 and 2025). From our perspective, the cypress trees are in their mature stage and no silvicultural interventions were implemented throughout the study period, so the structure of individual trees is relatively stable. Nevertheless, the time inconsistency may introduce uncertainties.

Beyond data constraints, the generalizability of the proposed framework can be further improved. Bark beetle-induced mortality in conifers typically progresses through green, yellow, red, and gray attack stages, accompanied by gradual crown discoloration over time. In this study, we successfully captured the infested cypress in the red and gray phases, respectively. In the next step, we will continue the monitoring and verification of our method on the identification of the second stage of infection, and also test on other forest pests and diseases, e.g., pine wood nematode-infested pines. In addition, future research could further import multispectral images from UAVs and MLs to an earlier and more precise identification of infection.

6. Conclusions

In this study, we presented a practical detection framework for cypress bark beetle infestation that combines a one-time airborne LiDAR acquisition for individual-tree delineation with repeated UAV RGB imagery for infestation monitoring. Based on different VIs and the threshold segmentation approach, infested trees were identified at the individual-tree level, and the effectiveness of the proposed index and the associated classification framework was quantitatively assessed. The main conclusions are summarized as follows:

(1)

A total of 1368 individual trees were identified using the LiDAR point cloud data and LSS method. Visual interpretation based on RGB imagery verified that the individual tree segmentation achieved a Precision of 82.51%, a Recall of 80.00%, and an F1-score of 81.24%, demonstrating that this method enables reliable automatic identification at the single-tree scale.

(2)

Four VIs (VARI, VDVI, RGI, and CIVE) were used to identify infested trees at the red attack phase using a μ ± 2σ threshold. Among them, RGI showed the best performance, achieving 100% Precision, 86.96% Recall, and an F1-score of 93.02%.

(3)

Based on the RGI index, a total of 20 insect-infested trees were identified, accounting for 1.46% of the cypress stand. The proposed method enables rapid and reliable monitoring of insect-infested trees, offering a critical tool for safeguarding forest health and advancing sustainable forest management.

(4)

In the supplementary experiment focusing on the gray attack stage, RGI again demonstrated superior performance compared with the other indices, confirming its robustness across two infestation phases. In addition, the μ ± 2σ thresholding approach showed better performance than the IQR-based method, indicating its methodological advantage within the proposed framework.