Individual Tree Detection Research Articles

Cross-Comparison of Individual Tree Detection Methods Using Low and High Pulse Density Airborne Laser Scanning Data

Numerous individual tree detection (ITD) methods have been developed for use with airborne laser scanning (ALS) data to provide tree-scale forest inventories across large spatial extents. Despite the growing number of methods, relatively few have been comparatively assessed using a single benchmark forest inventory validation dataset, limiting their operational application. In this study, we assessed seven ITD methods, representing three common approaches (point-cloud-based, raster-based, hybrid), across coniferous forest stands with diverse structure and composition to understand how ITD and height measurement accuracy vary with method, input parameters and data, and stand density. There was little variability in accuracy between the ITD methods where the average F-score and standard deviation (±SD) were 0.47 ± 0.03 using a lower pulse density ALS dataset with an average of 8 pulses per square meter (ppm2) and 0.50 ± 0.02 using a higher pulse density ALS dataset with an average of 22 ppm2. Using higher ALS pulse density data produced higher ITD accuracies (F-score increase of 10–13%) in some of the methods versus more modest gains in other methods (F-score increase of 1–3%). Omission errors were strongly related with stand density and largely consisted of suppressed trees underneath the dominant canopy. Simple canopy height model (CHM)-based methods that utilized fixed-size local maximum filters had the lowest omission errors for trees across all canopy positions. ITD accuracy had large intra-method variation depending on input parameters; however, the highest accuracies were obtained when parameters such as search window size and spacing thresholds were equal to or less than the average crown diameter of trees in the study area. All ITD methods produced height measurements for the detected trees that had low RMSE (<1.1 m) and bias (<0.5 m). Overall, the results from this study may help guide end-users with ITD method application and highlight future ITD method improvements.

Simulation-driven 3D forest growth forecasting based on airborne topographic LiDAR data and shading

Reliable forest growth forecasting requires detailed tree data for forest simulation, while manual on-site collection of relevant data is work-intensive and unfeasible in larger forests. This paper proposes a complete methodology for fully automated forest growth simulation that relies primarily on airborne topographic Light Detection And Ranging (LiDAR) point clouds of individual trees. The proposed method estimates tree parameters and performs growth of individual trees based on an individual-based forest growth simulator, named BWINPro. In addition, competition and detailed asymmetric tree crown growth are modeled regarding the shading of tree crowns, which is estimated from the surrounding environment and neighbor trees. The result of the proposed approach is a new point cloud for subsequent analyses. The proposed method was validated by comparing canopy height models derived from the point clouds of the simulated trees with canopy height models derived from more recent ground truth point clouds. The results demonstrate the efficacy of the proposed method which achieves a 9.4% higher accuracy than the averaged linear regression model and, in the case of datasets with more distinct self-standing trees, where a tree crown boundary plays major role, a 4.1% higher accuracy than the directly fitted linear regression model.

ReforesTree: A Dataset for Estimating Tropical Forest Carbon Stock with Deep Learning and Aerial Imagery

Forest biomass is a key influence for future climate, and the world urgently needs highly scalable financing schemes, such as carbon offsetting certifications, to protect and restore forests. Current manual forest carbon stock inventory methods of measuring single trees by hand are time, labour, and cost intensive and have been shown to be subjective. They can lead to substantial overestimation of the carbon stock and ultimately distrust in forest financing. The potential for impact and scale of leveraging advancements in machine learning and remote sensing technologies is promising, but needs to be of high quality in order to replace the current forest stock protocols for certifications. In this paper, we present ReforesTree, a benchmark dataset of forest carbon stock in six agro-forestry carbon offsetting sites in Ecuador. Furthermore, we show that a deep learning-based end-to-end model using individual tree detection from low cost RGB-only drone imagery is accurately estimating forest carbon stock within official carbon offsetting certification standards. Additionally, our baseline CNN model outperforms state-of-the-art satellite-based forest biomass and carbon stock estimates for this type of small-scale, tropical agro-forestry sites. We present this dataset to encourage machine learning research in this area to increase accountability and transparency of monitoring, verification and reporting (MVR) in carbon offsetting projects, as well as scaling global reforestation financing through accurate remote sensing.

Mapping the urban forest in detail: From LiDAR point clouds to 3D tree models

Trees are an integral component of the urban environment and important for human well-being, adaption measures to climate change and sustainable urban transformation. Understanding the small-scale impacts of urban trees and strategically managing the ecosystem services they provide requires high-resolution information on urban forest structure, which is still scarce. In contrast, there is an abundance of data portraying urban areas and an associated trend towards smart cities and digital twins as analysis platforms. A GIS workflow is presented in this paper that may close this data gap by classifying the urban forest from LiDAR point clouds, detecting and reconstructing individual crowns, and enabling a tree representation within semantic 3D city models. The workflow is designed to provide robust results for point clouds with a density of at least 4 pts/m2 that are widely available. Evaluation was conducted by mapping the urban forest of Dresden (Germany) using a point cloud with 4 pts/m². An object-based data fusion approach is implemented for the classification of the urban forest. A classification accuracy of 95 % for different urban settings is achieved by combining LiDAR with multispectral imagery and a 3D building model. Individual trees are detected by local maxima filtering and crowns are segmented using marker-controlled watershed segmentation. Evaluation highlights the influences of both urban and forest structure on individual tree detection. Substantial differences in detection accuracies are evident between trees along streets (72 %) and structurally more complex tree stands in green areas (31 %), as well as dependencies on tree height and crown diameter. Furthermore, an approach for parameterized reconstruction of tree crowns is presented, which enables efficient and realistic city-wide modeling. The suitability of LiDAR to measure individual tree metrics is illustrated as well as a framework for modeling individual tree crowns via geometric primitives.

A Conventional Cruise and Felled-Tree Validation of Individual Tree Diameter, Height and Volume Derived from Airborne Laser Scanning Data of a Loblolly Pine (P. taeda) Stand in Eastern Texas

Globally, remotely sensed data and, in particular, Airborne Laser Scanning (ALS), are being assessed by the forestry industry for their ability to acquire accurate forest inventories at an individual-tree level. This pilot study compares an inventory derived using the ForestView® biometrics analysis system to traditional cruise measurements and felled tree measurements for 139 Pinus taeda sp. (loblolly pine) trees in eastern Texas. The Individual Tree Detection (ITD) accuracy of ForestView® was 97.1%. In terms of tree height accuracy, ForestView® results had an overall lower mean bias and RMSE than the traditional cruise techniques when both datasets were compared to the felled tree data (LiDAR: mean bias = 1.1 cm, RMSE = 41.2 cm; Cruise: mean bias = 13.8 cm, RMSE = 57.5 cm). No significant difference in mean tree height was observed between the felled tree, cruise, and LiDAR measurements (p-value = 0.58). ForestView-derived DBH exhibited a −2.1 cm bias compared to felled-tree measurements. This study demonstrates the utility of this newly emerging ITD software as an approach to characterize forest structure on similar coniferous forests landscapes.

UTILIZING SINGLE PHOTON LASER SCANNING DATA FOR ESTIMATING INDIVIDUAL TREE ATTRIBUTES

Abstract. Mapping and monitoring forest resources require collection of spatially explicit and timely remote sensing (RS) data. Although field measurements are still important, the RS-based forest inventory helps mapping large areas to be cheaper, faster, less labor intensive, and spatially more explicit. The single-photon laser (SPL) scanning data has been exploited for different forestry applications but lacks deep examination in mapping individual trees as well as being compared with ordinary laser scanning (Linear-mode, LML) data and different individual tree detection (ITD) methods. Hence, this research focuses on applying and comparing two datasets (SPL and LML) for extracting attributes of individual trees by applying two tree crown segmentation methods (local maxima and watershed segmentation) on both datasets. The results were validated over 49 field measured plots of different species, located in southern boreal forest.The SPL yielded more accurate results for both tree density and height estimation. Watershed segmentation method yielded more accurate results for tree density and height estimation in both LML and SPL datasets. Tree density was underestimated by 4.7% (rRMSE: 32.3%) for all species. Comparing tree density estimation of different species, it was most accurate in deciduous plots (rBias: −9.5, rRMSE: 17.0%). Tree height estimation with SPL explained the variations of field-measured height very well (R2 = 0.93), and was reliably accurate, underestimated by 3.4% (rRMSE: 7.0%). The mean tree height estimation was most accurate in pine plots (rBias: 4.9%, rRMSE: 1.1%). In this research, SPL represented reliable and usable point cloud data for estimating tree height and density.

Extraction of individual trees based on Canopy Height Model to monitor the state of the forest

Active remotely sensed data can be used to perform a variety of forestry tasks including stand characterization, inventory, and management of forest and fire behavior modeling. The present work investigates the potential of Airborne Laser Scanning (ALS) derived methods applied in the deciduous forest by processing an individual tree detection (ITD) based on canopy Height model (CHM) and tree segmentation of larger-area point clouds. Different algorithms are tested and their performances are evaluated to show which of them can provide the most adequate number of trees compared with the ground truth. Tree scale information is used in order to determine stand age. The forest height, structure, and density are specified by applying individual tree Detection (ITD) to calculate some forest attributes such as stem volume, forest uniformity, and biomass estimation. The major aim of this post is to examine the state of the forest to monitor it in real-time. We assume that utilizing the LM algorithm, which was originally built for ITD from LiDAR data, trees should be automatically distinguished from the ALS-derived CHM with reasonable accuracy. As a result, the present research work studies the fixed treetop window size (FWS), fixed smoothing window size (SWS), and variable window (VW) effect on ITD performance (RMSE=3.4% and R=0.88). It is obvious, from the obtained results that smaller window sizes result in more trees. In fact, the smallest trees obscured by the largest trees containing the highest points in the neighborhood are often ignored by large windows. Crown delineation is also explored to extract the height of the trees, radius crown and, 3D coordinates and to compare them to those detected by a Low Bluetooth sensor “iBeacon.”.

Mapping Tree Canopy in Urban Environments Using Point Clouds from Airborne Laser Scanning and Street Level Imagery

Resilient cities incorporate a social, ecological, and technological systems perspective through their trees, both in urban and peri-urban forests and linear street trees, and help promote and understand the concept of ecosystem resilience. Urban tree inventories usually involve the collection of field data on the location, genus, species, crown shape and volume, diameter, height, and health status of these trees. In this work, we have developed a multi-stage methodology to update urban tree inventories in a fully automatic way, and we have applied it in the city of Pamplona (Spain). We have compared and combined two of the most common data sources for updating urban tree inventories: Airborne Laser Scanning (ALS) point clouds combined with aerial orthophotographs, and street-level imagery from Google Street View (GSV). Depending on the data source, different methodologies were used to identify the trees. In the first stage, the use of individual tree detection techniques in ALS point clouds was compared with the detection of objects (trees) on street level images using computer vision (CV) techniques. In both cases, a high success rate or recall (number of true positive with respect to all detectable trees) was obtained, where between 85.07% and 86.42% of the trees were well-identified, although many false positives (FPs) or trees that did not exist or that had been confused with other objects were always identified. In order to reduce these errors or FPs, a second stage was designed, where FP debugging was performed through two methodologies: (a) based on the automatic checking of all possible trees with street level images, and (b) through a machine learning binary classification model trained with spectral data from orthophotographs. After this second stage, the recall decreased to about 75% (between 71.43 and 78.18 depending on the procedure used) but most of the false positives were eliminated. The results obtained with both data sources were robust and accurate. We can conclude that the results obtained with the different methodologies are very similar, where the main difference resides in the access to the starting information. While the use of street-level images only allows for the detection of trees growing in trafficable streets and is a source of information that is usually paid for, the use of ALS and aerial orthophotographs allows for the location of trees anywhere in the city, including public and private parks and gardens, and in many countries, these data are freely available.

Optimizing aerial imagery collection and processing parameters for drone‐based individual tree mapping in structurally complex conifer forests

Abstract Recent advances in remotely piloted aerial systems (‘drones’) and imagery processing enable individual tree mapping in forests across broad areas with low‐cost equipment and minimal ground‐based data collection. One such method involves collecting many partially overlapping aerial photos, processing them using ‘structure from motion’ (SfM) photogrammetry to create a digital 3D representation and using the 3D model to detect individual trees. SfM‐based forest mapping involves myriad decisions surrounding methods and parameters for imagery acquisition and processing, but it is unclear how these individual decisions or their combinations impact the quality of the resulting forest inventories. We collected and processed drone imagery of a moderate‐density, structurally complex mixed‐conifer stand. We tested 22 imagery collection methods (altering flight altitude, camera pitch and image overlap), 12 imagery processing parameterizations (image resolutions and depth map filtering intensities) and 286 tree detection methods (algorithms and their parameterizations) to create 7,568 tree maps. We compared these maps to a 3.23‐ha ground reference map of 1,775 trees >5 m tall that we created using traditional field survey methods. The accuracy of individual tree detection (ITD) and the resulting tree maps was generally maximized by collecting imagery at high altitude (120 m) with at least 90% image‐to‐image overlap, photogrammetrically processing images into a canopy height model (CHM) with a twofold upscaling (coarsening) step and detecting trees from the CHM using a variable window filter after applying a moving window mean smooth to the CHM. Using this combination of methods, we mapped trees with an accuracy exceeding expectations for structurally complex forests (for canopy‐dominant trees >10 m tall, sensitivity = 0.69 and precision = 0.90). Remotely measured tree heights corresponded to ground‐measured heights with R2 = 0.95. Accuracy was higher for taller trees and lower for understorey trees and would likely be higher in less dense and less structurally complex stands. Our results may guide others wishing to efficiently produce broad‐extent individual tree maps of conifer forests without investing substantial time tailoring imagery acquisition and processing parameters. The resulting tree maps create opportunities for addressing previously intractable ecological questions and informing forest management.

Correcting the Results of CHM-Based Individual Tree Detection Algorithms to Improve Their Accuracy and Reliability

Individual tree detection algorithms (ITD) are used to obtain accurate information about trees. Following the process of individual tree detection, it is possible to use additional processing tools to determine tree parameters such as tree height, crown base height, crown volume, or stem volume. However, many of the methods developed so far have focused on parameterising the algorithms based on the study area, height structure or tree species analysed. Applying the parameters of the method can be challenging in areas with dense and heterogeneous forests with a diverse stand structure. Therefore, this work aimed to develop a method to correct the results of ITD algorithms to identify individual trees more reliably, taking into account different ITD methods based on the Canopy Height Model. In the present study, we proposed a three-step approach to correct segmentation errors. In the first step, erroneous (under- and over-segmentation errors) and correct segments were classified. After classification, the second step was to refine the under-segmentation errors. The final step was to merge segments from the over-segmentation class with correct segments based on the specified conditions. The study was conducted in one of the most complex and diverse forest communities in Europe, making tree identification a major challenge. The accuracy of the segmentation improvements varied depending on the method applied and tree species group examined. Thus, based on the results, the paper advocates for the correction method due to its efficiency in mixed forest stands. Therefore, the present study offers a possible solution to reduce segmentation errors by considering different forest types and different CHM-based ITD methods for identifying individual trees.

Upscaling aboveground biomass of larch (Larix olgensis Henry) plantations from field to satellite measurements: a comparison of individual tree-based and area-based approaches

ABSTRACT Currently, aboveground biomass (AGB) estimation primarily relies on an area-based approach (ABA), particularly at a large scale. With the advancement of individual tree detection techniques and the availability of multi-platform remotely sensed data, the individual tree-based approach (ITA) provides the potential of accurate and nondestructive fine AGB mapping. However, the performances of the two approaches on upscaling AGB from individual tree to stand level have not been thoroughly investigated. This study conducted the nondestructive AGB estimation using both ITA and ABA and compared the AGB estimates and their spatial uncertainties from an individual tree- to stand-level based on multi-platform LiDAR and Landsat 8 OLI imagery, taking Larix olgensis that is one of the most notable afforestation tree species in northeastern China as an example. Results showed that the point cloud segmentation (PCS) outperformed CHM-based individual tree crown delineation algorithms and obtained the highest accuracy of individual tree AGB estimate (R2 = 0.97, RMSE = 28.58 kg, rRMSE = 21.13%) for the dense larch plantations (about 1265 trees/ha). The plot-level AGB estimate aggregated by all detected trees and its uncertainty based on Monte Carlo simulation were 158.16 and 3.64 Mg·ha−1, respectively. The average pixel-level AGB of larch plantations in Maorshan Forest Farm estimated by ITA and ABA were similar (129.66 v.s 144.38 Mg·ha−1). ITA outperformed ABA in terms of pixel-level AGB accuracy and spatial uncertainty of pixel-level and stand-level AGB estimates. The overestimation of low AGB values, typical in ABA, was effectively eliminated by the ITA. The upscaling frameworks proposed in this study for AGB estimation and spatial uncertainty quantification based on ITA and ABA could be extended to other plantations or uncomplex forests. This study contributes to the accurate quantification of AGB and understanding of the uncertainties in the carbon stock of forest ecosystems at multi-scales.

Individual tree detection and estimation of stem attributes with mobile laser scanning along boreal forest roads

The collection of field-reference data is a key task in remote sensing-based forest inventories. However, traditional methods of collection demand extensive personnel resources. Thus, field-reference data collection would benefit from more automated methods. In this study, we proposed a method for individual tree detection (ITD) and stem attribute estimation based on a car-mounted mobile laser scanner (MLS) operating along forest roads. We assessed its performance in six ranges with increasing mean distance from the roadside. We used a Riegl VUX-1LR sensor operating with high repetition rate, thus providing detailed cross sections of the stems. The algorithm we propose was designed for this sensor configuration, identifying the cross sections (or arcs) in the point cloud and aggregating those into single trees. Furthermore, we estimated diameter at breast height (DBH), stem profiles, and stem volume for each detected tree. The accuracy of ITD, DBH, and stem volume estimates varied with the trees’ distance from the road. In general, the proximity to the sensor of branches 0–10 m from the road caused commission errors in ITD and over estimation of stem attributes in this zone. At 50–60 m from roadside, stems were often occluded by branches, causing omissions and underestimation of stem attributes in this area. ITD’s precision and sensitivity varied from 82.8% to 100% and 62.7% to 96.7%, respectively. The RMSE of DBH estimates ranged from 1.81 cm (6.38%) to 4.84 cm (16.9%). Stem volume estimates had RMSEs ranging from 0.0800 m3 (10.1%) to 0.190 m3 (25.7%), depending on the distance to the sensor. The average proportion of detected reference volume was highly affected by the performance of ITD in the different zones. This proportion was highest from 0 to 10 m (113%), a zone that concentrated most ITD commission errors, and lowest from 50 to 60 m (66.6%), mostly due to the omission errors in this area. In the other zones, the RMSE ranged from 87.5% to 98.5%. These accuracies are in line with those obtained by other state-of-the-art MLS and terrestrial laser scanner (TLS) methods. The car-mounted MLS system used has the potential to collect data efficiently in large-scale inventories, being able to scan approximately 80 ha of forests per day depending on the survey setup. This data collection method could be used to increase the amount of field-reference data available in remote sensing-based forest inventories, improve models for area-based estimations, and support precision forestry development.

Individual Tree Detection in Urban ALS Point Clouds with 3D Convolutional Networks

Since trees are a vital part of urban green infrastructure, automatic mapping of individual urban trees is becoming increasingly important for city management and planning. Although deep-learning-based object detection networks are the state-of-the-art in computer vision, their adaptation to individual tree detection in urban areas has scarcely been studied. Some existing works have employed 2D object detection networks for this purpose. However, these have used three-dimensional information only in the form of projected feature maps. In contrast, we exploited the full 3D potential of airborne laser scanning (ALS) point clouds by using a 3D neural network for individual tree detection. Specifically, a sparse convolutional network was used for 3D feature extraction, feeding both semantic segmentation and circular object detection outputs, which were combined for further increased accuracy. We demonstrate the capability of our approach on an urban topographic ALS point cloud with 10,864 hand-labeled ground truth trees. Our method achieved an average precision of 83% regarding the common 0.5 intersection over union criterion. 85% percent of the stems were found correctly with a precision of 88%, while tree area was covered by the individual tree detections with an F1 accuracy of 92%. Thereby, we outperformed traditional delineation baselines and recent detection networks.

Image to attribute model for trees (ITAM-T): individual tree detection and classification in Alberta boreal forest for wildland fire fuel characterization

ABSTRACT Regional and municipal decision makers rely on fuel (vegetation) maps to inform decisions on tree stand management related to wildfire management and response. Remote sensing of trees is used in commercial applications but has limited uptake in the fire management community. A two-stage detection and identification convolutional network for high-resolution RGB drone imagery is developed to address this limitation. The detection routine is based on DeepForest, an existing convolutional neural network implementation designed to recognize trees in aerial imagery. Retraining the model and implementing an adaptive window-size workflow improves tree detection, with F1 scores reaching 85% and averaging 72% for k-fold cross-validation in boreal forest. For classification, a VGG19 network with added data augmentation and dropout layers is trained. When this network is implemented, manually annotated trees are recognized as coniferous with an average F1 of 97% and deciduous with an 87% F1. Overall, the developed image-to-attribute model for trees reaches a maximum F1 score of 85% considering classification after identification, with averages of 72% for coniferous trees and 57% for deciduous trees over six sites. Tree height, size, and stem density are extracted from the tree location output and geometric data. The calculated density is compared to the density of manual annotations, with an average R2 of 0.90. A remote preliminary proximity-based hazard assessment is performed on a rural property in Alberta, demonstrating the model’s ability to detect and classify trees near values-at-risk. The results indicate a potential extension to low-cost decision support in enterprise and fire-related applications.

Predicting Individual Tree Diameter of Larch (Larix olgensis) from UAV-LiDAR Data Using Six Different Algorithms

Individual tree detection is an increasing trend in LiDAR-based forest inventories. The locations, heights, and crown areas of the detected trees can be estimated rather directly from the LiDAR data by using the LiDAR-based canopy height model and segmentation methods to delineate the tree crowns. However, the most important tree variable is the diameter of the tree stem at the breast height (DBH) which can seldom be interpreted directly from the LiDAR data. Therefore, the use of individually detected trees in forest planning calculations requires predictions for the DBH. This study tested six methods for predicting the DBH from laser scanning data collected by an unmanned aerial vehicle from Larix olgensis plantations located in northeast China. The tested methods were the linear regression model (LM), a linear model with ridge regularization (LMR), support vector regression (SVR), random forest (RF), artificial neural network (ANN), and the k-nearest neighbors (KNN) method. Both tree-level and stand-level metrics derived from the LiDAR point cloud data (for instance percentiles of the height distribution of the echoes) were used as potential predictors of DBH. Compared to the LM, all other methods improved the accuracy of the predictions. On the other hand, all methods tended to underestimate the DBH of the largest trees, which could be due to the inability of the methods to sufficiently describe nonlinear relationships unless different transformations of the LiDAR metrics are used as predictors. The support vector regression was evaluated to be the best method for predicting individual tree diameters from LiDAR data. The benefits of the methods tested in this study can be expected to be the highest in the case of little prior knowledge on the relationships between the predicted variable and predictors, a high number of potential predictors, and strong mutual correlations among the potential predictors.

High-Density UAV-LiDAR in an Integrated Crop-Livestock-Forest System: Sampling Forest Inventory or Forest Inventory Based on Individual Tree Detection (ITD)

Lidar point clouds have been frequently used in forest inventories. The higher point density has provided better representation of trees in forest plantations. So we developed a new approach to fill this gap in the integrated crop-livestock-forest system, the sampling forest inventory, which uses the principles of individual tree detection applied under different plot arrangements. We use a UAV-lidar system (GatorEye) to scan an integrated crop-livestock-forest system with Eucalyptus benthamii seed forest plantations. On the high density UAV-lidar point cloud (>1400 pts. m2), we perform a comparison of two forest inventory approaches: Sampling Forest Inventory (SFI) with circular (1380 m2 and 2300 m2) and linear (15 trees and 25 trees) plots and Individual Tree Detection (ITD). The parametric population values came from the approach with measurements taken in the field, called forest inventory (FI). Basal area and volume estimates were performed considering the field heights and the heights measured in the LiDAR point clouds. We performed a comparison of the variables number of trees, basal area, and volume per hectare. The variables by scenarios were submitted to analysis of variance to verify if the averages are considered different or equivalent. The RMSE (%) were calculated to explain the deviation between the measured volume (filed) and estimated volume (LiDAR) values of these variables. Additionally, we calculated rRMSE, Standard error, AIC, R2, Bias, and residual charts. The basal area values ranged from 7.40 m2 ha−1 (C1380) to 8.14 m2 ha−1 281 (C2300), about −5.9% less than the real value (8.65 m2 ha−1). The C2300 scenario was the only one whose confidence interval (CI) limits included the basal area real. For the total stand volume, the ITD scenario was the one that presented the closer values (689.29 m3) to the real total value (683.88 m3) with the real value positioned in the CI. Our findings indicate that for the stand conditions under study, the SFI approach (C2300) that considers an area of 2300 m2 is adequate to generate estimates at the same level as the ITD approach. Thus, our study should be able to assist in the selection of an optimal plot size to generate estimates with minimized errors and gain in processing time.

A Voxel-Based Individual Tree Stem Detection Method Using Airborne LiDAR in Mature Northeastern U.S. Forests

This paper describes a new method for detecting individual tree stems that was designed to perform well in the challenging hardwood-dominated, mixed-species forests common to the northeastern U.S., where canopy height-based methods have proven unreliable. Most prior research in individual tree detection has been performed in homogenous coniferous or conifer-dominated forests with limited hardwood presence. The study area in central Pennsylvania, United States, includes 17+ tree species and contains over 90% hardwoods. Existing methods have shown reduced performance as the proportion of hardwood species increases, due in large part to the crown-focused approaches they have employed. Top-down approaches are not reliable in deciduous stands due to the inherent complexity of the canopy and tree crowns in such stands. This complexity makes it difficult to segment trees and accurately predict tree stem locations based on detected crown segments. The proposed voxel column-based approach has advantages over both traditional canopy height model-based methods and computationally demanding point-based solutions. The method was tested on 1125 reference trees, ≥10 cm diameter at breast height (DBH), and it detected 68% of all reference trees and 87% of medium and large (sawtimber-sized) trees ≥28 cm DBH. Significantly, the commission rate (false predictions) was negligible as most raw false positives were confirmed in follow-up field visits to be either small trees below the threshold for recording or trees that were otherwise missed during the initial ground survey. Minimizing false positives was a priority in tuning the method. Follow-up in-situ evaluation of individual omission and commission instances was facilitated by the high spatial accuracy of predicted tree locations generated by the method. The mean and maximum predicted-to-reference tree distances were 0.59 m and 2.99 m, respectively, with over 80% of matches within <1 m. A new tree-matching method utilizing linear integer programming is presented that enables rigorous, repeatable matching of predicted and reference trees and performance evaluation. Results indicate this new tree detection method has potential to be operationalized for both traditional forest management activities and in providing the more frequent and scalable inventories required by a growing forest carbon offsets industry.

Comparison of Classical Methods and Mask R-CNN for Automatic Tree Detection and Mapping Using UAV Imagery

Detecting and mapping individual trees accurately and automatically from remote sensing images is of great significance for precision forest management. Many algorithms, including classical methods and deep learning techniques, have been developed and applied for tree crown detection from remote sensing images. However, few studies have evaluated the accuracy of different individual tree detection (ITD) algorithms and their data and processing requirements. This study explored the accuracy of ITD using local maxima (LM) algorithm, marker-controlled watershed segmentation (MCWS), and Mask Region-based Convolutional Neural Networks (Mask R-CNN) in a young plantation forest with different test images. Manually delineated tree crowns from UAV imagery were used for accuracy assessment of the three methods, followed by an evaluation of the data processing and application requirements for three methods to detect individual trees. Overall, Mask R-CNN can best use the information in multi-band input images for detecting individual trees. The results showed that the Mask R-CNN model with the multi-band combination produced higher accuracy than the model with a single-band image, and the RGB band combination achieved the highest accuracy for ITD (F1 score = 94.68%). Moreover, the Mask R-CNN models with multi-band images are capable of providing higher accuracies for ITD than the LM and MCWS algorithms. The LM algorithm and MCWS algorithm also achieved promising accuracies for ITD when the canopy height model (CHM) was used as the test image (F1 score = 87.86% for LM algorithm, F1 score = 85.92% for MCWS algorithm). The LM and MCWS algorithms are easy to use and lower computer computational requirements, but they are unable to identify tree species and are limited by algorithm parameters, which need to be adjusted for each classification. It is highlighted that the application of deep learning with its end-to-end-learning approach is very efficient and capable of deriving the information from multi-layer images, but an additional training set is needed for model training, robust computer resources are required, and a large number of accurate training samples are necessary. This study provides valuable information for forestry practitioners to select an optimal approach for detecting individual trees.

Detection of Individual Trees in UAV LiDAR Point Clouds Using a Deep Learning Framework Based on Multichannel Representation

Individual tree detection is critical for forest investigation and monitoring. Several existing methods have difficulties to detect trees in complex forest environments due to insufficiently mining descriptive features. This study proposes a deep learning (DL) framework based on a designed multichannel information complementarity representation for detecting trees in complex forest using UAV laser scanning point clouds. The proposed method consists of two main stages: ground filtering and tree detection. In the first stage, a modified graph convolution network with a local topological information layer is designed to separate the ground points. Unlike most existing parametric methods, our ground filtering method avoids the optimal parameters selection to adapt to different kinds of environments. For tree detection, a top-down slice (TDS) module is first designed to mine the vertical structure information in a top-down way. Then, a special multichannel representation (MCR) is developed to preserve different distribution patterns of points from complementary perspectives. Finally, a multibranch network (MBNet) is proposed for individual tree detection by fusing multichannel features, which can provide discriminative information for MBNet to detect trees more accurately. MBNet was evaluated on seven forest areas [UAV light detection and ranging (LiDAR) data with the mean size of 14 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$000~\text {m}^{2}$ </tex-math></inline-formula> and point density of 250 points/ <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text {m}^{2}$ </tex-math></inline-formula> ]. Experimental results showed that the proposed framework achieves excellent performance. Our method obtains promising performance with a mean recall of 89.23% and a mean F1-score of 87.04%.

Recognition of specified errors of Individual Tree Detection methods based on Canopy Height Model

In the last two decades, many individual tree detection (ITD) methods have been developed based on Airborne Laser Scanner (ALS) data, but their application still has some shortcomings. In complex and dense forest stands, the number of tree segments extracted from Canopy Height Model (CHM) are frequently over- and under-estimated, reducing the ability to predict stand structure. This study presents the possibility of distinguishing correct from erroneous segments resulting from the ITD methods based on CHM. For this purpose, three machine learning methods were tested: Random Forest (RF), Support Vector Machine and k-Nearest Neighbor among which the RF algorithm gave the best results. Groups of predictors based on segment geometry as well as structural and intensity metrics from the ALS point cloud were used. From the whole set of predictors, the ratio of segment perimeter to its area proved to be the most important. Using RF classifier, it was possible to identify under-segmentation and over-segmentation errors, as well as correct segments, with high accuracies for training (OA = 87.0% and κ = 0.794) and test data (OA = 85.3 and κ = 0.641). Recognition of specific segmentation errors is important in that it can be used to determine which conditions favour the occurrence of errors and which tree species are more likely to be incorrectly segmented. Therefore, the study verified the error susceptibility of individual tree species. This can help improve calibration and development of segmentation algorithms according to the tree species analysed.