Individual tree-level structural parameters are crucial for sustainable forest management and the evaluation of forest growth, as well as their ecological functions. The advanced Unmanned Aerial Vehicle-based Laser Scanning (ULS) and Backpack Laser Scanning (BLS) techniques offer a detailed three-dimensional insight into the forest structure information, have great potential for accurately estimating forest structural parameters in forest research and practices. However, tree-level key structural parameters (e.g., DBH, height and volume, etc.) extracted from ULS and BLS point clouds were usually evaluated by field measurement which has relatively low accuracy and high uncertainties especially under the complicated forest conditions. To quantify and comprehensively analyze the performance of ULS and BLS point cloud data in extracting key structural parameters of individual trees, this study conducted a multi-platform laser scanning remote sensing campaign in a typical Chinese fir plantation at Yangkou State Forest Farm, Fujian Province, along with field measurements of 16 sample Chinese fir trees. We processed the ULS and BLS point clouds and extracted structural parameters at treelevel by advanced algorithms, and then evaluated the tree height, DBH and volume by measurements from the field as well as felled trees through tree stem analysis by fitting a variety of taper models. Results demonstrated that 1) ULS outperforms ultrasonic altimeters in quantifying individual tree heights compared to felled tree height. The RMSE of tree height estimates from manually measured ULS data and automatically extracted ULS data were 0.77 m (4.00%) and 0.87 m (4.53%), respectively. 2) The accuracy of DBH estimated from BLS data was relatively high compared to field measured data, with an R2 of 0.96 (RMSE = 1.98 cm, relative RMSE (rRMSE) = 8.37%). 3) The Yamamoto-style binary volume model fitted in this study performed the best among all the models, with an R2 of 0.95 (RMSE = 0.05069 m3, rRMSE = 14.90%), followed by the binary volume model derived from the integral operation of the taper equation and the existing local generalized binary volume model. This study collected precious data by tree stem analysis through felled trees, also provided a solid quantitative demonstration of the robust capabilities of ULS in tree height estimation and BLS in DBH estimation, and their combination in estimating volume, which has rarely been done in previous studies.

The high-precision instance segmentation of tree saplings is a fundamental prerequisite for the high-throughput phenotypic analysis of individual seedlings in intelligent tree breeding and precision silviculture. However, sapling segmentation remains challenging because of blurred boundaries, object adhesion, missed detections, and inaccurate mask delineation in field environments. To improve sapling segmentation performance and address these challenges, this study proposes a multimodal Mask R-CNN framework in which RGB imagery was paired with one multispectral-derived vegetation index at a time to construct separate RGB-VI input combinations, taking ginkgo saplings as a representative case. A dataset of 400 saplings was constructed using a high-throughput field phenotyping platform. The backbone network was extended with an independent vegetation index branch, and three fusion strategies (early, multi-step, and late fusion) were designed within a feature pyramid network to enable multi-scale multimodal feature integration. The results showed that all multimodal models outperformed unimodal baselines in terms of segmentation accuracy and recall. Among them, the multi-step fusion strategy achieved the best performance, while the RGB-EVI multi-step fusion model achieved the highest strict-matching precision (AP@75 = 87.7%) and recall (71.3%), with superior performance in dense sapling delineation and background suppression. These findings indicate that multimodal feature fusion can effectively improve sapling instance segmentation and provide methodological support for high-throughput plant phenotyping.

Canopy Structural Complexity (CSC) is a key structural attribute for evaluating forest ecological functions and health, with accurate canopy extraction serving as the prerequisite for its quantification and the basis for improving metric reliability and elucidating spatiotemporal canopy dynamics. However, existing canopy extraction methods generally rely on large-scale fully labeled point cloud datasets, which are often impractical for large-scale Unmanned Aerial Vehicle (UAV) LiDAR forest surveys. To address this limitation, we propose a weakly supervised canopy extraction strategy that relies on a limited number of labeled plots. By incorporating a pseudo-label generation strategy and a data-augmentation-based consistency self-supervised constraint, the method effectively improves segmentation performance in unlabeled scenarios. Furthermore, we introduce a multi-dimensional joint canopy entropy index that integrates both global and local canopy features, combining projection-based global canopy entropy with graph-based local canopy entropy to provide a more comprehensive CSC quantification. Experiments were conducted using three labeled plots as the training set and a held-out labeled plot for quantitative evaluation, with numerous unlabeled plots used for qualitative generalization analysis. Results showed that on the labeled test plots, the proposed weakly supervised strategy achieved 85.37% mIoU and 92.64% OA, surpassing the best-performing fully supervised method. Visual comparisons in real, unlabeled forest scenes demonstrate that the proposed method maintains clear boundaries between canopy and non-canopy regions, even in areas with significant topographic variation, fragmented canopy, or complex structures, while effectively suppressing false positives in ground and low-vegetation areas. The proposed multi-dimensional joint canopy entropy, when applied to the predictions of the weakly supervised canopy extraction strategy, successfully quantified the CSC using large-scale, unlabeled UAV forest LiDAR point clouds. It yielded results consistent with ecological priors and exhibited robustness under varying levels of point cloud sparsification. Based on normalized scores across three evaluation metrics, the proposed index achieved a score of 0.76, outperforming several commonly used CSC quantification indices.

Solar-induced chlorophyll fluorescence (SIF) is a promising photosynthetic proxy and a valuable phenotyping tool. However, under drought and fluctuating light, the mechanistic linkages between SIF and photosynthesis require further investigation due to complex light energy partitioning. In this study, we measured leaf-level SIF, gas-exchange, and active fluorescence parameters in Ginkgo biloba under drought and fluctuating light. Our results demonstrate a drought-induced decoupling of both leaf-level SIF and PSII-emitted chlorophyll fluorescence (SIFPSII, derived from active fluorescence parameters) from photosynthesis. This was driven by a slight SIFPSII reduction alongside strong photosynthetic suppression. The limited SIFPSII decrease resulted from a stable fluorescence yield (ΦSIF), maintained by the regulatory interplay between non-photochemical quenching (NPQ) and the fraction of open PSII reaction centers (qL), and the low responsiveness of ΦSIF to changes in the maximum photochemical quantum yield (ΦPSIImax). In contrast, photosynthetic suppression was largely due to decreased stomatal conductance (gs). ΦPSIImax and photorespiration are also central to balancing the light and carbon reactions, preventing energy overload. Interestingly, fluctuating light phases did not significantly alter this decoupling. However, when combining data from both high and low light, SIFPSII and photosynthesis remained tightly coupled, likely due to their shared light dependency. These insights improve the physiological interpretation of SIF and underscore the importance of SIF-based multi-trait phenotyping models for accurate field monitoring.

The estimation of leaf nitrogen content (LNC) is important for assessing the growth status of Ginkgo saplings. UAV hyperspectral imagery provides an effective data source for estimating LNC. However, canopy structural variation can affect hyperspectral signals. Although canopy scattering coefficient (CSC) correction can reduce such effects, extracting nitrogen-sensitive spectral features remains challenging. In this study, we applied CSC correction and continuous wavelet transform (CWT) to UAV hyperspectral imagery. Six models were first evaluated to determine the optimal modeling strategy after CSC correction. CWT was then performed at five scales, and the top 10% nitrogen-sensitive wavelet coefficients were selected according to their correlation with LNC. Based on the selected coefficients, we constructed and optimized three types of wavelet indices using an exhaustive search strategy. We then compared the optimized index with existing reflectance-based vegetation indices. The results showed that the optimized wavelet index achieved the best LNC prediction performance (R2=0.691, RMSE=1.856 g·kg⁻¹). This study highlights the potential of combining CSC correction with CWT for UAV-based LNC estimation in Ginkgo saplings.

Intelligent forest tree breeding has advanced plant phenotyping, yet existing research largely focuses on large-leaf agricultural crops, with limited attention to fine-grained leaf analysis of sapling trees in open-field environments. Natural scenes introduce challenges including scale variation, illumination changes, and irregular leaf morphology. To address these issues, we collected UAV RGB imagery of field-grown saplings and constructed the Poplar-leaf dataset, containing 1,202 branches and 19,876 pixel-level annotated leaf instances. To our knowledge, this is the first instance segmentation dataset specifically designed for forestry leaves in open-field conditions. We propose LeafInst, a novel segmentation framework tailored for irregular and multi-scale leaf structures. The model integrates an Asymptotic Feature Pyramid Network (AFPN) for multi-scale perception, a Dynamic Asymmetric Spatial Perception (DASP) module for irregular shape modeling, and a dual-residual Dynamic Anomalous Regression Head (DARH) with Top-down Concatenation decoder Feature Fusion (TCFU) to improve detection and segmentation performance. On Poplar-leaf, LeafInst achieves 68.4 mAP, outperforming YOLOv11 by 7.1 percent and MaskDINO by 6.5 percent. On the public PhenoBench benchmark, it reaches 52.7 box mAP, exceeding MaskDINO by 3.4 percent. Additional experiments demonstrate strong generalization and practical utility for large-scale leaf phenotyping.

Accurate detection of urban vegetation changes is essential for maintaining ecological balance and promoting sustainable urban development. Deep learning has become a leading technology for urban vegetation change detection, owing to its superior performance in feature representation and detection accuracy. High-resolution RGB remote sensing imagery has emerged as the key data source for deep learning-based urban vegetation change detection, due to its fine spatial detail and lower acquisition cost compared to high-resolution multispectral imagery. However, most of the existing urban vegetation change detection methods based on deep learning are supervised and require manually annotated training samples, limiting their automation and efficiency. In this study, we proposed a novel unsupervised urban vegetation change detection approach (CCST-KFN) for high-resolution RGB remote sensing imagery. Specifically, an advanced cross-scale change sample transfer (CCST) method was developed to automatically generate reliable high-resolution training samples. It first produces medium-resolution urban vegetation changed and unchanged samples by pre-detection, leveraging the more discriminative spectral features available in medium-resolution multispectral imagery. These samples are then mapped to the high-resolution domain guided by change information. In addition, a dedicated urban vegetation change detection network by fusing domain knowledge and change features (KFN) was designed for urban vegetation change mapping. It takes as input two high-resolution RGB images and two resampled enhanced normalized difference vegetation index images, which emphasize urban vegetation by combining near-infrared, red, and blue bands. This enables the network to effectively exploit the comprehensive features and domain knowledge of urban vegetation changes. Experiments on three typical urban datasets demonstrate that the proposed CCST-KFN identified urban vegetation changes more completely and with fewer errors compared to seven unsupervised change detection methods. Moreover, its application to a representative urban area in Nanjing City, China, further validates its practical effectiveness. Overall, the proposed CCST-KFN offers a reliable and automated solution for accurate urban vegetation change detection using high-resolution RGB remote sensing imagery, with the potential to further support urban ecological security and sustainable development.

Individual tree-level fine-scale (especially stem-and branch-scales) structural parameters constitute a critical foundation for tree structural trait assessments, biomass component estimations and tree physiological property evaluations. However, the Unmanned Aerial Vehicle (UAV) Laser Scanning (ULS) has limitations in sampling distance and penetrating capability in dense tree canopies, thus restricting its ability to extract detailed stem-and branch-scale structural parameters. The emergence of both advanced ULS and Backpack Laser Scanning (BLS) technologies have potential to precisely extract fine-scale structural parameters of individual trees. In this study, we proposed an advanced ALE-CS-NGMS approach for individual tree (Poplar (Populus spp.)) stem-and branch-scale structural parameters extraction by ULS and BLS point clouds. First, an Adaptive Least-squares Ellipse (ALE) fitting algorithm was developed to accurately derive the stem diameter of individual trees. Second, a Canopy-stem Separation (CS) model was built by identifying canopy point cloud through derivatives based on the vertical distribution profile of individual trees, while canopy volume was delineated by the AlphaShape as well as a voxel-based algorithm. Finally, a method integrating Neighborhood Graphs and Minimum Spanning (NGMS) was developed to extract individual tree stem, and stem taper curves were fitted to estimate individual-tree stem volume. The results demonstrated that the developed ALE approach yielded a root mean square error (RMSE) of 2.87 cm, representing an accuracy enhancement approximately 0.47 cm for DBH estimation. The NGMS approach produced RMSEs of 0.33 m(3) and 0.40 m(3) for stem volume estimation by using BLS and BLS + ULS data. The CS model achieved RMSEs of 6.48 m(3) and 3.48 m(3) for canopy volume estimation with the BLS and BLS + ULS data, respectively. Branch inclination angles exhibited an increase with stand age, generally ranging between 60 degrees and 100 degrees. The distribution of branch inclination across stands of varying ages revealed that in the 8year-old and 12-year-old plots, branch angles fell within the 60 degrees-90 degrees interval.

Accurate estimation of chlorophyll content helps understand the physiological health of saplings. Traditional field measurement methods limit large-scale applications, while remote sensing monitoring methods, though improving efficiency, still face challenges in data analysis and processing. A key challenge is extracting the canopy of individual trees and estimating chlorophyll content amid interference from weed vegetation. To overcome this challenge, this study develops a novel coupled deep learning approach. First, an Attention and Pyramid U-Net (APUNet) was proposed to improve the segmentation accuracy of Ginkgo seedlings crown boundaries in multispectral drone images. Second, the hyper-parameter combination of the You Only Look Once (YOLO) model was optimized to improve sapling detection accuracy under complex conditions with high weed coverage. Next, a sliding window method was developed to reduce duplicate detection rates for detecting all saplings in the entire image, and the intersection of APUNet and detection results was used for segmentation of individual saplings. Finally, a PROSAIL-Deep Neural Network (DNN) model for chlorophyll estimation was developed, and the impact of using APUNet YOLO to remove weed pixels on chlorophyll estimation was investigated, along with how over-coverage and under-coverage during the segmentation process affect the accuracy of chlorophyll estimation. The results show that, compared to U-Net, APUNet improved segmentation performance, with the IoU increasing from 0.697 to 0.736. After hyper-parameter optimization through grid search, the YOLOv8l model achieved an mAP50 of 0.968 and when this model was applied to full-image detection using a sliding window approach, its duplicate detection rate (ddr) decreased by 11.2% compared with the general method. In chlorophyll inversion, the chlorophyll estimation accuracy reached its highest after using APUNet-YOLO to reduce the interference from weed pixels, with an R2 of 0.752 and an RMSE of 7.143 mu g/ cm2. Meanwhile, the sensitivity analysis of the impact of segmentation errors on chlorophyll estimation accuracy shows that, compared to under-coverage segmentation, over-inclusion segmentation leads to a greater decline in chlorophyll estimation accuracy. This novel coupled model can effectively monitor the chlorophyll content of Ginkgo saplings in settings with complex backgrounds, providing a practical possibility for the realization of precision forestry.

The accurate and efficient extraction of individual tree phenotypic traits for seabuckthorn (Hippophae rhamnoides L.) in natural forests is crucial for germplasm exploration, precision silviculture, and ecological restoration. This study extracted structural and biochemical traits of seabuckthorn in Tibet’s Lhasa valley using Unmanned aerial vehicle (UAV) LiDAR, multispectral imagery, and the N-PROSAIL model. Firstly, building on a classification conducted through multi-scale spatial analysis and hierarchical clustering with dynamic thresholds, shrub interference was effectively reduced, thereby improving the accuracy of individual tree segmentation. Tree height and crown width were derived from the segmentation results, and a DBH estimation model was developed using handheld LiDAR data. Finally, leaf nitrogen content was mapped within canopies using random forest combined with the N-PROSAIL model and nitrogen reference data. The results demonstrated that the optimized segmentation method successfully extracted structural traits (F1 = 84.21%). Tree height was accurately estimated (R2 = 0.814, RMSE = 0.580 m), and the DBH prediction model performed satisfactorily (R2 = 0.779, RMSE = 1.725 cm). The random forest model also effectively estimated leaf nitrogen content (R2 = 0.680, RMSE = 2.074 mg/g).

Tree-level structural parameters estimation plays a key role in the researches and practice in sustainable forest management, carbon storage estimation, as well as ecological function evaluation. However, single Light Detection and Ranging (LiDAR) platform exhibits limitations when acquiring complete (i.e., including over-story and under-story) point cloud data for forest stands, e.g., UAV LiDAR systems tend to overlook details of the tree trunk or the lower ground, while Backpack LiDAR systems struggle to capture the treetop, etc. The limited shared features of point clouds from UAV and Backpack LiDAR sensors also pose challenges in the accurate registration and merging of these datasets. In this study, we proposed a marker free automatic registration framework for multi-platform forest point clouds with terrain features. The framework comprised three key stages: first, a curvature-adaptive weighting mechanism was adapted to optimized the Fast Point Feature Histogram (FPFH) descriptors for initial coarse registration, utilizing terrains features. Second, individual tree positions were extracted from each platform's LiDAR dataset and employed as key feature points for matching. Third, a similarity function was constructed to evaluate the most geometrically consistent point correspondences across platforms, which were subsequently refined through an Iterative Closest Point (ICP) algorithm. Furthermore, a voxel-based denoising algorithm that integrated point density with vertical connectivity was developed to identify and filter out noise from the backpack LiDAR data-specifically, non-structural elements such as branches and shrubs. This denoising process laid a robust foundation for accurately locating individual tree centers. Additionally, a layer-wise adaptive circular fitting method was introduced for determining trunk positions. By clustering trunk point clouds at successive vertical layers, this method yielded precise estimates of straight, individual tree trunk centers for use in subsequent registration steps. The proposed framework achieved a registration accuracy of RMSE = 0.098-0.134 m across diverse forest types and terrain conditions, demonstrating its robustness and applicability in complex environments. This facilitated the integration of UAV and backpack LiDAR technologies in forestry resource monitoring. Using the fused point cloud data, tree-level structural parameters estimation of diameter at breast height (RMSE = 1-1.2 cm), tree height (RMSE = 0.29-0.55 m).

Individual tree structure and wood density are important indicators of forest quality and key parameters for biomass calculation. To explore the extraction accuracy of individual tree structure parameters based on LiDAR technology, as well as the correlation between individual tree structure parameters, resistance value and wood density can be beneficial for providing new ideas for predicting wood density. Taking a 23-year-old Ginkgo plantation as the research object, the tree QSM (Quantitative Structure Model) was constructed based on terrestrial and backpack LiDAR point clouds, and the individual tree structure parameters were extracted. The accuracy of estimating structure parameters based on two types of point clouds was compared. A wood density prediction model was constructed using principal component analysis based on the resistance, diameter, tree height, and crown width. The accuracy verification was carried out and it showed that the estimation accuracies of individual tree structure parameters (DBH, tree height, and crown width) extracted from tree QSM constructed based on TLS and BLS all had R2 > 0.8. The estimation accuracy of DBH based on TLS was slightly higher than that based on BLS, and the estimation accuracy of tree height and crown width based on TLS was slightly lower than that based on BLS. BLS has great potential in accurately obtaining forest structure information, improving forest information collection efficiency, promoting forest resource monitoring, forest carbon sink estimation, and forest ecological research. The feasibility of predicting the wood basic density based on wood resistance (R2 = 0.51) and combined with DBH, tree height, and crown width (R2 = 0.49) was relatively high. Accurate and non-destructive estimation of the wood characteristics of standing timber can guide forest cultivation and management and promote sustainable management and utilization of forests.

Since nitrogen (N) underpins plant vitality by forming proteins, nucleic acids, and chlorophyll, quantifying it through leaf nitrogen concentration (LNC, %) becomes pivotal for growth assessment and precision forestry. However, the three-dimensional structure of the canopy, crown shadow and other background factors complicate the estimation of LNC from crown bidirectional reflectance factor (BRF). To address these challenges, we employ canopy scattering coefficients (CSC) to analyze light behavior within canopies. Accurate estimation of N relies on the association of N with chlorophyll, dry matter, water and canopy structure. To improve the LNC prediction, we developed an enhanced spectral index model called the Difference combined Simple Ratio index (DSR), which improves LNC estimation by minimizing the effects of canopy structure and shadows. Results indicate that the shadow-filtering index methods effectively eliminated shadowed pixels, enhancing the correlation between single-band reflectance and LNC. The CSC-based DSR is the best crown-level LNC estimation for Liriodendron sino-americanum plantation (R-2 > 0.77, RMSE < 0.7). The estimation model exhibits significant potential for mapping crown-scale LNC distributions in Liriodendron sino-americanum plantations, as well as improving the understanding of the confounding effects of canopy structure and shadows on the LNC estimation.

The identification of individual trees can reveal the competitive and symbiotic relationships among trees within forest stands, which is fundamental understand biodiversity and forest ecosystems. Highly precise identification of individual trees can significantly improve the efficiency of forest resource inventory, and is valuable for biomass measurement and forest carbon storage assessment. In previous studies through deep learning approaches for identifying individual tree, feature extraction is usually difficult to adapt to the variation of tree crown architecture, and the loss of feature information in the multi-scale fusion process is also a marked challenge for extracting trees by remote sensing images. Based on the one-stage deep learning network structure, this study improves and optimizes the three stages of feature extraction, feature fusion and feature identification in deep learning methods, and constructs a novel feature-oriented individual tree identification network (FO-Net) suitable for UAV high-resolution images. Firstly, an adaptive feature extraction algorithm based on variable position drift convolution was proposed, which improved the feature extraction ability for the individual tree with various crown size and shape in UAV images. Secondly, to enhance the network's ability to fuse multiscale forest features, a feature fusion algorithm based on the "gather-and-distribute" mechanism is proposed in the feature pyramid network, which realizes the lossless cross-layer transmission of feature map information. Finally, in the stage of individual tree identification, a unified self-attention identification head is introduced to enhanced FO-Net's perception ability to identify the trees with small crown diameters. FO-Net achieved the best performance in quantitative analysis experiments on self-constructed datasets, with mAP50, F1-score, Precision, and Recall of 90.7%, 0.85, 85.8%, and 82.8%, respectively, realizing a relatively high accuracy for individual tree identification compared to the traditional deep learning methods. The proposed feature extraction and fusion algorithms have improved the accuracy of individual tree identification by 1.1% and 2.7% respectively. The qualitative experiments based on Grad-CAM heat maps also demonstrate that FO-Net can focus more on the contours of an individual tree in high-resolution images, and reduce the influence of background factors during feature extraction and individual tree identification. FO-Net deep learning network improves the accuracy of individual trees identification in UAV high-resolution images without significantly increasing the parameters of the network, which provides a reliable method to support various tasks in fine-scale precision forestry.

The precise extractions of tree components such as wood (i.e., trunk and branches) and leaves are fundamental prerequisites for obtaining the key attributes of trees, which will provide significant benefits for ecological and physiological studies and forest applications. Terrestrial laser scanning technology offers an efficient means for acquiring three-dimensional information on tree attributes, and has marked potential for extracting the detailed tree attributes of tree components. However, previous studies on wood–leaf separation exhibited limitations in unsupervised adaptability and robustness to complex tree architectures, while demonstrating inadequate performance in fine branch detection. This study proposes a novel unsupervised model (NE-PC) that synergizes geometric features with graph-based path analysis to achieve accurate wood–leaf classification without training samples or empirical parameter tuning. First, the boundary-preserved supervoxel segmentation (BPSS) algorithm was adapted to generate supervoxels for calculating geometric features and representative points for constructing the undirected graph. Second, a node expansion (NE) approach was proposed, with nodes with similar curvature and verticality expanded into wood nodes to avoid the omission of trunk points in path frequency detection. Third, a path concatenation (PC) approach was developed, which involves detecting salient features of nodes along the same path to improve the detection of tiny branches that are often deficient during path retracing. Tested on multi-station TLS point clouds from trees with complex leaf–branch architectures, the NE-PC model achieved a 94.1% mean accuracy and a 86.7% kappa coefficient, outperforming renowned TLSeparation and LeWos (ΔOA = 2.0–29.7%, Δkappa = 6.2–53.5%). Moreover, the NE-PC model was verified in two other study areas (Plot B, Plot C), which exhibited more complex and divergent branch structure types. It achieved classification accuracies exceeding 90% (Plot B: 92.8 ± 2.3%; Plot C: 94.4 ± 0.7%) along with average kappa coefficients above 80% (Plot B: 81.3 ± 4.2%; Plot C: 81.8 ± 3.2%), demonstrating robust performance across various tree structural complexities.

As an essential part of terrestrial ecosystems, forests are key to sustaining ecological balance, supporting the carbon cycle, and offering various ecosystem services. In recent years, forests in Southwest China have experienced notable greening. However, the rising occurrence and severity of droughts present a significant threat to the stability of forest ecosystems in this region. This study adopted the near-infrared reflectance of vegetation (NIRv) and the lag-1 autocorrelation of NIRv as indicators to assess the dynamics and resilience of forests in Southwest China. We identified a progressive decline in forest resilience since 2008 despite a dominant greening trend in Southwest China’s forests during the last 20 years. By developing the eXtreme Gradient Boosting (XGBoost) model and Shapley additive explanation framework (SHAP), we classified forests in Southwest China into coniferous and broadleaf types to evaluate the driving factors influencing changes in forest resilience and mapped the spatial distribution of dominant drivers. The results showed that the resilience of coniferous forests was mainly driven by variations in elevation and land surface temperature (LST), with mean absolute SHAP values of 0.045 and 0.038, respectively. In contrast, the resilience of broadleaf forests was primarily influenced by changes in photosynthetically active radiation (PAR) and soil moisture (SM), with mean absolute SHAP values of 0.032 and 0.028, respectively. Regions where elevation and LST were identified as dominant drivers were mainly distributed in coniferous forest areas across central, eastern, and northern Yunnan Province as well as western Sichuan Province, accounting for 32.9% and 20.0% of the coniferous forest area, respectively. Meanwhile, areas where PAR and SM were dominant drivers were mainly located in broadleaf forest regions in Sichuan and eastern Guizhou, accounting for 29.9% and 27.7% of the broadleaf forest area, respectively. Our study revealed that the forest greening does not necessarily accompany an enhancement in resilience in Southwest China, identifying the driving factors behind the decline in forest resilience and highlighting the necessity of differentiated restoration strategies for forest ecosystems in this region.

Bamboo forests are natural habitat for the giant panda which is one of the most vulnerable mammal species. In structurally complex natural forests, bamboos are normally located under the canopy of taller trees, which makes them difficult to be quantified accurately. Although Light Detection and Ranging (LiDAR) technologies have been well established as the effective tool for forest structure assessment, the use of LiDAR to assess understory bamboo in structurally complex natural forests is less well known. We present a novel vertical vegetation classification (VVC) approach to map the structure of understory bamboos for giant panda forage in natural forests. An optimized demarcation point identification (DPI) model was developed for stratifying different vertical layers from coarse to fine scales. Three-dimensional understory bamboo point clouds were successfully isolated from the forest point cloud, then bamboo structure predictive models were developed through understory bamboo point cloud metrics and applied over the entire study area to generate spatially continuous maps of understory bamboo structure. Our results indicate that the isolation of the understory bamboo point cloud using the developed VVC approach performs well and has small bias, the extracted maximum height is close to fieldmeasured maximum height (R2 = 0.77, rRMSE = 15.02 %). Height-related metrics have higher correlations with bamboo structure (mean natural and true height, basal diameter, and total aboveground biomass) than other metrics (r > 0.8), and understory bamboo structures are estimated with relatively high accuracy (R2 = 0.84 - 0.91, rRMSE = 10.87 - 29.41 %). We also find varying effects of topography on the spatial distribution of different understory bamboo species. This study demonstrates the benefits of utilizing LiDAR data to ascertain fine-scale understory bamboo resources, providing critical supports for giant panda habitat assessment and conservation.

Precise measurement of individual tree aboveground biomass (AGB) is essential for effective plantation management and tree cultivars development. However, the complex structural and species mixture of Liriodendron sino-americanum forests pose challenges on accurate AGB estimation. This study developed an advanced approach for extracting individual Liriodendron sino-americanum trees from mixed forests using LiDAR-derived Canopy Height Model (CHM) metrics and high-resolution imagery, UAV and backpack LiDAR data were registered using seed points neighborhood features and Euclidean distance, the registered seed points of the backpack LiDAR were used for individual tree segmentation in UAV LiDAR, and AGB was then calculated at multi-scale (i. e., tree-and plot-level) by the synergetic implementations of UAV-based metrics, backpack LiDAR-derived diameter at breast height (DBH) (the DBH was extracted by developing a novel algorithm to integrate continuum elements, normal vector angles, KD Tree smoothing and least-squares circle fitting) and the fitted models. Results showed that 1) UAV and backpack LiDAR data matching based on seed points was effective, with an RMSE of 9.72 cm, and improved the UAV LiDAR segmentation accuracy (from 0.76 to 0.87). 2) The novel DBH extraction algorithm achieved an R2 of 0.89 and an RMSE of 1.78 cm. 3) UAV LiDAR canopy volume metric accurately estimated AGB for individual Liriodendron sino-americanum trees, with an R2 of 0.91, RMSE of 8.38 kg. Additionally, the LAI_pc (from LiDAR) metric was more suitable for predicting AGB in pure forests, while the D6 (canopy return density) metric provided superior estimates for intermediate proportions. This integrated method demonstrates high potential for precise, scalable AGB estimation in structurally complex and mixed-species forests, contributing to enhanced forest inventory, carbon accounting, and ecological research.

Plant phenomics, the comprehensive study of plant phenotypes, has gained prominence as a vital tool for understanding the intricate relationships between genotypes and the environment. Image-based plant phenomics has progressed rapidly, and three-dimensional (3D) phenotyping is a valuable extension of traditional 2D phenomics. However, the increased data dimensionality poses challenges to feature extraction and phenotyping. In recent decades, deep learning has led to remarkable progress in revolutionizing 3D phenotyping. Therefore, this review highlights the importance of using deep learning in 3D plant phenomics. It systematically overviews the capabilities of deep learning for 3D computer vision, covering 3D representation, classification, detection and tracking, semantic segmentation, instance segmentation, and generation. Additionally, deep learning techniques for 3D point preprocessing (e.g., annotation, downsampling, and dataset organization) and various plant phenotyping tasks are discussed. Finally, the challenges and perspectives associated with deep learning in 3D plant phenomics are summarized, including (1) benchmark dataset construction by using synthetic datasets and methods such as generative artificial intelligence and unsupervised or weakly supervised learning; (2) accurate and efficient 3D point cloud analysis by leveraging multitask learning, lightweight models, and self-supervised learning; and (3) deep learning for 3D plant phenomics by exploring interpretability, extensibility, and multimodal data utilization. The exploration of deep learning in 3D plant phenomics is poised to spur breakthroughs in a new dimension of plant science.

Ecological niche models (ENMs) are crucial for identifying habitat distribution patterns, understanding habitat preferences, and formulating effective conservation policies. However, accurately quantifying the three-dimensional (3D) structure of habitats, a fundamental component, presents challenges. These estimations heavily depend on the quality of original samples (presence/absence), yet reliable absence data requires prolonged and repeated observations, limiting both efficiency and accuracy. In the study, we focused on the endangered Yunnan snub-nosed monkey (Rhinopithecus bieti), listed on the International Union for Conservation of Nature (IUCN) Red List. We developed an adaptive similarity-based model that introduced a "similarity" pseudo-absence sampling approach for ecological niche modeling using fine-scale (20 m) 3D environmental variables from UAV LiDAR data. This approach integrated geographic similarity with an adaptive kernel density estimation (AKDE) method to prioritize pseudo-absence data sampling and then employed three typical machine learning models (SVM, BRT, and RF) for prediction, verifying the feasibility of this approach and offering direct insights into habitat distribution and preferences. The results indicated that the AKDE method provided the best fit in measuring similarity features. Through the model, the performance of estimations exhibited improved (AUC = 0.89-0.94, TSS = 0.71-0.82, and COR = 0.69-0.79), with average increases of 7 %, 14 %, and 12 %, respectively. The RF model produced more coherent suitable habitats, identifying regions at higher elevations (3100 m-3300 m) with preferences for low understory vegetation density (2-5 m, <10 %), moderate canopy relief ratio (> 0.35), lower tree height (10 m-25 m), and sunny slopes (0.60-1). Our findings demonstrate that integrating UAV LiDAR data with ecological niche modeling, along with the improved pseudo-absence sampling approach, enhances habitat assessment and offers significant potential for advancing conservation strategies.