Tree Detection Algorithm Research Articles

Cross-site learning in deep learning RGB tree crown detection

Tree crown detection is a fundamental task in remote sensing for forestry and ecosystem ecology. While many individual tree segmentation algorithms have been proposed, the development and testing of these algorithms is typically site specific, with few methods evaluated against data from multiple forest types simultaneously. This makes it difficult to determine the generalization of proposed approaches, and limits tree detection at broad scales. Using data from the National Ecological Observatory Network, we extend a recently developed deep learning approach to include data from a range of forest types to determine whether information from one forest can be used for tree detection in other forests, and explore the potential for building a universal tree detection algorithm. We find that the deep learning approach works well for overstory tree detection across forest conditions. Performance was best in open oak woodlands and worst in alpine forests. When models were fit to one forest type and used to predict another, performance generally decreased, with better performance when forests were more similar in structure. However, when models were pretrained on data from other sites and then fine-tuned using a relatively small amount of hand-labeled data from the evaluation site, they performed similarly to local site models. Most importantly, a model fit to data from all sites performed as well or better than individual models trained for each local site.

Quantification of uncertainty in aboveground biomass estimates derived from small-footprint airborne LiDAR

To address uncertainty in biomass estimates across spatial scales, we determined aboveground biomass (AGB) in Californian forests through the use of individual tree detection methods applied to small-footprint airborne LiDAR. We propagated errors originating from a generalized allometric equation, LiDAR measurements, and individual tree detection algorithms to AGB estimates at the tree and plot levels. Larger uncertainties than previously reported at both tree and plot levels were found when AGB was derived from remote sensing. On average, per-tree AGB error was 135% of the estimated AGB, and per-plot error was 214% of the estimated AGB. We found that from tree to plot level, the allometric equation constituted the largest proportion of the total AGB uncertainty. The proportion of the uncertainty associated with remote sensing errors was larger in lower AGB forests, and it decreased as AGB increased. The framework in which we performed the error propagation analysis can be used to address AGB uncertainties in other ecosystems and can be integrated with other analytical techniques.

Urban tree health assessment using airborne hyperspectral and LiDAR imagery

Urban trees provide valuable ecosystem services but are at the same time under continuous pressure due to unfavorable site conditions. In order to better protect and manage our natural capital, urban green managers require frequent and detailed information on tree health at the city wide scale. In this paper we developed a workflow to monitor tree defoliation and discoloration of broadleaved trees in Brussels, Belgium, through the combined use of airborne hyperspectral and LiDAR data. Individual trees were delineated using an object-based tree detection and segmentation algorithm primarily based on LiDAR data with an average accuracy of 91%. We constructed Partial Least Squares Regression (PLSR) models to derive tree chlorophyll content (RMSE = 2.8 μg/cm²; R² = 0.77) and Leaf Area Index (LAI; RMSE = 0.5; R² = 0.66) from the average canopy spectrum. Existing spectral indices were found to perform significantly worse (RMSE > 7 μg/cm² and >1.5 respectively), mainly due to contamination of tree spectra by neighboring background materials. In the absence of local calibration data, the applicability of PLSR to other areas, sensors and tree species might be limited. Therefore, we identified the best performing/least sensitive spectral indices and proposed a simple pixel selection procedure to reduce disturbing background effects. For LAI, laser penetration metrics derived from LiDAR data attained comparable accuracies as PLSR and were suggested instead. Detection of healthy and unhealthy trees based on remotely sensed tree properties matched reasonably well with a more traditional visual tree assessment (93% and 71% respectively). If combined with early tree stress detection methods, the proposed methodology would constitute a solid basis for future urban tree health monitoring programs.

Efficiency of Individual Tree Detection Approaches Based on Light-Weight and Low-Cost UAS Imagery in Australian Savannas

The reliability of airborne light detection and ranging (LiDAR) for delineating individual trees and estimating aboveground biomass (AGB) has been proven in a diverse range of ecosystems, but can be difficult and costly to commission. Point clouds derived from structure from motion (SfM) matching techniques obtained from unmanned aerial systems (UAS) could be a feasible low-cost alternative to airborne LiDAR scanning for canopy parameter retrieval. This study assesses the extent to which SfM three-dimensional (3D) point clouds—obtained from a light-weight mini-UAS quadcopter with an inexpensive consumer action GoPro camera—can efficiently and effectively detect individual trees, measure tree heights, and provide AGB estimates in Australian tropical savannas. Two well-established canopy maxima and watershed segmentation tree detection algorithms were tested on canopy height models (CHM) derived from SfM imagery. The influence of CHM spatial resolution on tree detection accuracy was analysed, and the results were validated against existing high-resolution airborne LiDAR data. We found that the canopy maxima and watershed segmentation routines produced similar tree detection rates (~70%) for dominant and co-dominant trees, but yielded low detection rates (<35%) for suppressed and small trees due to poor representativeness in point clouds and overstory occlusion. Although airborne LiDAR provides higher tree detection rates and more accurate estimates of tree heights, we found SfM image matching to be an adequate low-cost alternative for the detection of dominant and co-dominant tree stands.

マルチロータヘリを用いた森林内樹木の計測

Japan is a forestry country where forests account for about 70% of the land, and introduction of robots is expected in the field of forestry. However, there is already a method of measuring height using aircraft and laser scanner, but the diameter of breast height is investigated by human hands despite the lack of personnel. Therefore, in this research, we propose a trees diameter measurement system using multi rotor helicopter. We aim to stably fly through the forest and measure the breast height diameter of trees in the forest. We propose a tree detection algorithm using LRF data.

Estimation of 2-D Clutter Maps in Complex Under-Canopy Environments From Airborne Discrete-Return Lidar

Detection of near-ground objects occluded by above-ground vegetation from airborne light detection and ranging (lidar) measurements remains challenging. Our hypothesis is that the probability of obstruction due to objects above ground at any location in the forest environment can be reasonably characterized solely from airborne lidar data. The essence of our approach is to develop a data-driven learning scheme that creates high-resolution two-dimensional (2-D) probability maps for obstruction in the under-canopy environment. These maps contain information about the probabilities of obstruction (clutter map) and lidar undersampling (uncertainty map) in the near-ground space. Airborne and terrestrial lidar data and field survey data collected within the forested mountainous environment of Shenandoah National Park, Virginia, USA are utilized to test and evaluate the proposed approach in this work. A newly developed individual tree detection algorithm is implemented to estimate the undersampled stem contributions to the probability of obstruction. Results show the effectiveness of the tree detection algorithm with an accuracy index (AI) of between 61.5% and 80.7% (tested using field surveys). The estimated clutter maps are compared to the maps created from terrestrial scans (i.e., ground truth) and the results show the root-mean-square error (RMSE) of 0.28, 0.32, and 0.34 at three study sites. The overall framework in deriving near-ground clutter and uncertainty maps from airborne lidar data would be useful information for the prediction of line-of-sight visibility, mobility, and above-ground forest biomass.

Development of an automated individual tree detection model using point cloud LiDAR data for accurate tree counts in a Pinus radiata plantation

ABSTRACTWe develop and evaluate a new individual tree detection (ITD) algorithm to automatically locate and estimate the number of individual trees within a Pinus radiata plantation from relatively sparse airborne LiDAR point cloud data. The area of interest comprised stands covering a range of age classes and stocking levels. Our approach is based on local maxima (LM) filtering that tackles the issue of selecting the optimal search radius from the LiDAR point cloud for every potential LM using metrics derived from local neighbourhood data points; thus, it adapts to the local conditions, irrespective of canopy variability. This was achieved through two steps: (i) logistic regression model development using simulated stands composed of individual trees derived from real LiDAR point cloud data and (ii) application testing of the model using real plantation LiDAR point cloud data and geolocated, tree-level reference crowns that were manually identified in the LiDAR imagery. Our ITD algorithm performed well compared with previous studies, producing RMSE of 5.7% and a bias of only −2.4%. Finally, we suggest that the ITD algorithm can be used for accurately estimating stocking and tree mapping, which in turn could be used to derive the plot-level metrics for an area-based approach for enhancing estimates of stand-level inventory attributes based on plot imputation.

Weak Connection Edges Independent Discriminant of Rapid Spanning Tree Recommendation of Social Network Community

Aiming at the problem of low accuracy and high computational complexity of the traditional social network community recommendation algorithm, a rapid spanning tree detection algorithm is proposed to independently discriminate social network community with the weak connected edge, in order to improve the accuracy of community recommendation and reduce the complexity of algorithm. Firstly, according to the characteristics of social network community recommendation, the maximum spanning tree algorithm is proposed, which is based on the edge weight distribution node similarity, to realize the effective detection of social network community. Secondly, for the proposed algorithm having the problems of repeated adding and deleting of weakly connected edges and the waste of computing resources, a rapid spanning tree detection algorithm based on the independent discrimination of weakly connected edge is proposed so as to further improve the calculation efficiency of the algorithm. Lastly, the effectiveness of the proposed algorithm is verified by comparing the experimental results in the standard test database.

Calibrated tree counting on remotely sensed images of planted forests

The development of robust and accurate methods for counting trees from remotely sensed data could provide substantial cost savings in forest inventory. A new methodology that provides a framework for calibrating tree detection algorithms to obtain accurate tree counts for even-aged stands is described. The methodology was evaluated using two tree detection algorithms and two operators using airborne laser scanning (ALS) and orthophotograph images for four Pinus radiata D.Don stands ranging in age between 5 and 32 years with stand densities ranging between 204 and 826 stems ha−1. For application of the methodology to ALS images the error of estimate on the total count was 4.7% when calibration counts from actual ground plots were used and 10.5% when calibration counts from virtual plots on the image were used. For orthophotographs, the error of estimate was 6.1% using ground calibration plots and 24.3% using calibration counts from virtual plots. The described methodology was shown to be robust to variations in the process from the two operators and two algorithms evaluated. The measure of accuracy determined using the methodology can be used to provide an objective basis for evaluating a wide range of tree counting and detection processes in future research.

Using Tree Detection Algorithms to Predict Stand Sapwood Area, Basal Area and Stocking Density in Eucalyptus regnans Forest

Managers of forested water supply catchments require efficient and accurate methods to quantify changes in forest water use due to changes in forest structure and density after disturbance. Using Light Detection and Ranging (LiDAR) data with as few as 0.9 pulses m−2, we applied a local maximum filtering (LMF) method and normalised cut (NCut) algorithm to predict stocking density (SDen) of a 69-year-old Eucalyptus regnans forest comprising 251 plots with resolution of the order of 0.04 ha. Using the NCut method we predicted basal area (BAHa) per hectare and sapwood area (SAHa) per hectare, a well-established proxy for transpiration. Sapwood area was also indirectly estimated with allometric relationships dependent on LiDAR derived SDen and BAHa using a computationally efficient procedure. The individual tree detection (ITD) rates for the LMF and NCut methods respectively had 72% and 68% of stems correctly identified, 25% and 20% of stems missed, and 2% and 12% of stems over-segmented. The significantly higher computational requirement of the NCut algorithm makes the LMF method more suitable for predicting SDen across large forested areas. Using NCut derived ITD segments, observed versus predicted stand BAHa had R2 ranging from 0.70 to 0.98 across six catchments, whereas a generalised parsimonious model applied to all sites used the portion of hits greater than 37 m in height (PH37) to explain 68% of BAHa. For extrapolating one ha resolution SAHa estimates across large forested catchments, we found that directly relating SAHa to NCut derived LiDAR indices (R2 = 0.56) was slightly more accurate but computationally more demanding than indirect estimates of SAHa using allometric relationships consisting of BAHa (R2 = 0.50) or a sapwood perimeter index, defined as (BAHaSDen)½ (R2 = 0.48).

Generalized Reduced-State Vector Sequence Detection

A multi-dimensional set-partitioning scheme is proposed for joint reduced-state sequence detection (JRSSD) of vector sequences corrupted by intersymbol interference (ISI) and additive white Gaussian noise (AWGN). The existing multi-dimensional set-partitioning schemes result in complexity exponential in the dimensionality of the vectors in the vector sequence. The proposed set-partitioning scheme allows the construction of reduced-state trellises with very few states even when the vectors are of high dimensionality. One of the key requirements of the JRSSD algorithm is an efficient algorithm for spatial multi-symbol detection, for which we propose the reduced-state tree (RST) detection algorithm. In addition to performance improvement, the inclusion of RST algorithm for spatial multi-symbol detection provides a common framework for designing receivers for both flat-fading and frequency-selective fading MIMO channels.

Efficient lossless codec for still color images with backward coding of wavelet trees

Despite the approval of a new standard in 2009, JPEG-extended range lossless, current digital products still employ previous standards for lossless image compression, such as JPEG, JPEG2000, JPEG-LS, etc. Wavelet-based codecs can provide abundant functionalities and excellent compression efficiency. Among them, the backward coding of wavelet trees (BCWT) algorithm offers lower complexity and consumes less internal buffer memory without sacrificing quality at similar compression ratios (CR) when compared to other wavelet-based codecs, such as JPEG2000 and set partitioning in hierarchical trees (SPIHT). A line-based BCWT was developed for further reduction of internal buffer memory. A very efficient line-based lossless BCWT compression algorithm is presented. Lossless color and lossless wavelet transform are employed and the original BCWT algorithm is modified for lossless operation, including incorporation of adaptive arithmetic coding. In order to eliminate coding redundancies, a set to zeros method and a zero tree detection algorithm are proposed, which significantly enhance the boundary condition CR performance while reserving the algorithm’s advantages. Tests and analysis results show that the lossless BCWT algorithm requires less memory and computational resources than SPIHT and JPEG2000, while retaining image quality comparable to the standard image codecs, therefore, lossless BCWT is quite suitable for implementation in modern digital technologies.

An Accuracy Assessment of Tree Detection Algorithms in Juniper Woodlands

This research provides a comprehensive accuracy assessment of five methods for classifying western juniper (Juniperus occidentalis) canopy cover from 1 m, 4-band National Agriculture Imagery Program (NAIP) imagery. Two object-oriented classification approaches (image segmentation and spatial wavelet analysis, (SWA)) are compared to three pixel based classification approaches (random forests, Iterative Self-Organizing Data Analysis (ISODATA), and maximum likelihood). Methods are applied to approximately 250 km 2 in the intermountain western USA. A robust suite of statistical approaches, which offer an alternative to traditional kappa-based methods, are utilized to determine equivalence between methods and overall effectiveness. Object-oriented approaches have the highest overall accuracy among the assessed methods. Each of the methods varied considerably in cover class accuracy. SWA has the highest class accuracy when juniper canopy cover is low (0 to 40 percent cover), ISODATA performs best at moderate cover (60 to 80 percent) and maximum likelihood performs best at higher cover (60 to 100 percent cover).

Enhanced detection of 3D individual trees in forested areas using airborne full-waveform LiDAR data by combining normalized cuts with spatial density clustering

Abstract. A detailed understanding of the spatial distribution of forest understory is important but difficult. LiDAR remote sensing has been developing as a promising additional instrument to the conventional field work towards automated forest inventory. Unfortunately, understory (up to 50% of the top-tree height) in mixed and multilayered forests is often ignored due to a difficult observation scenario and limitation of the tree detection algorithm. Currently, the full-waveform (FWF) LiDAR with high penetration ability against overstory crowns can give us new hope to resolve the forest understory. Former approach based on 3D segmentation confirmed that the tree detection rates in both middle and lower forest layers are still low. Therefore, detecting sub-dominant and suppressed trees cannot be regarded as fully solved. In this work, we aim to improve the performance of the FWF laser scanner for the mapping of forest understory. The paper is to develop an enhanced methodology for detecting 3D individual trees by partitioning point clouds of airborne LiDAR. After extracting 3D coordinates of the laser beam echoes, the pulse intensity and width by waveform decomposition, the newly developed approach resolves 3D single trees are by an integrated approach, which delineates tree crowns by applying normalized cuts segmentation to the graph structure of local dense modes in point clouds constructed by mean shift clustering. In the context of our strategy, the mean shift clusters approximate primitives of (sub) single trees in LiDAR data and allow to define more significant features to reflect geometric and reflectional characteristics towards the single tree level. The developed methodology can be regarded as an object-based point cloud analysis approach for tree detection and is applied to datasets captured with the Riegl LMS-Q560 laser scanner at a point density of 25 points/m2 in the Bavarian Forest National Park, Germany, respectively under leaf-on and leaf-off conditions. The experiments lead to a detection rate of up to 67% for trees in the middle height layer and up to 53% for trees in the lower forest layer. It corresponds to an overall improvement in the detection rate of nearly 25% for forest understory compared to that obtained by the former method by extracting individual trees using normalized cuts segmentation solely.

Efficient Soft-Input Soft-Output Tree Detection via an Improved Path Metric

Tree detection techniques are often used to reduce the complexity of a posteriori probability (APP) detection in multiantenna wireless communication systems. In this paper, we introduce an efficient soft-input soft-output tree detection algorithm that employs a new type of look-ahead path metric in the process of branch pruning (or sorting). While conventional path metrics depend only on symbols on a visited path, the new path metric accounts for unvisited parts of the tree in advance through an unconstrained linear estimator and adds a bias term that reflects the contribution of as-yet undecided symbols. By applying the linear estimate-based look-ahead path metric to an -algorithm that selects the best paths for each level of the tree, we develop a new soft-input soft-output tree detector, called an improved soft-input soft-output -algorithm (ISS-MA). Based on an analysis of the probability of correct path loss, we show that the improved path metric offers substantial performance gain over the conventional path metric. We also demonstrate through simulations that the proposed ISS-MA can be a promising candidate for soft-input soft-output detection in high-dimensional systems.

Estimating Biophysical Parameters of Individual Trees in an Urban Environment Using Small Footprint Discrete-Return Imaging Lidar

Quantification of biophysical parameters of urban trees is important for urban planning, and for assessing carbon sequestration and ecosystem services. Airborne lidar has been used extensively in recent years to estimate biophysical parameters of trees in forested ecosystems. However, similar studies are largely lacking for individual trees in urban landscapes. Prediction models to estimate biophysical parameters such as height, crown area, diameter at breast height, and biomass for over two thousand individual trees were developed using best subsets multiple linear regression for a study area in central Oklahoma, USA using point cloud distributional metrics from an Optech ALTM 2050 lidar system. A high level of accuracy was attained for estimating individual tree height (R2 = 0.89), dbh (R2 = 0.82), crown diameter (R2 = 0.90), and biomass (R2 = 0.67) using lidar-based metrics for pooled data of all tree species. More variance was explained in species-specific estimates of biomass (R2 = 0.68 for Juniperus virginiana to 0.84 for Ulmus parviflora) than in estimates from broadleaf deciduous (R2 = 0.63) and coniferous (R2 = 0.45) taxonomic groups—or the data set analysed as a whole (R2 = 0.67). The metric crown area performed particularly well for most of the species-specific biomass equations, which suggests that tree crowns should be delineated accurately, whether manually or using automatic individual tree detection algorithms, to obtain a good estimation of biomass using lidar-based metrics.

Comparison of six individual tree crown detection algorithms evaluated under varying forest conditions

In this article, six individual tree crown (ITC) detection/delineation algorithms are evaluated, using an image data set containing six diverse forest types at different geographical locations in three European countries. The algorithms use fundamentally different techniques, including local maxima detection, valley following (VF), region-growing (RG), template matching (TM), scale-space (SS) theory and techniques based on stochastic frameworks. The structurally complexity of the forests in the aerial images used ranges from a homogeneous plantation and an area with isolated tree crowns to an extremely dense deciduous forest type. None of the algorithms alone could successfully analyse all different cases. The study shows that it is important to partition the imagery into homogeneous forest stands prior to the application of individual tree detection algorithms. It furthermore suggests a need for a common, publicly available suite of test images and common test procedures for evaluation of individual tree detection/delineation algorithms. Finally, it highlights that, for complex forest types, monoscopic images are insufficient for consistent tree crown detection, even by human interpreters.

Assessment of charcoal driven deforestation rates in a fragile rangeland environment in North Eastern Somalia using very high resolution imagery

Multi-temporal very high resolution satellite images and field work have been used for quantifying the tree cutting rate over a 5 years period in a very arid tiger bush area of North Eastern Somalia with intensive charcoal production activities. By applying both a classical area frame sampling approach with visual interpretation of the samples and a semi-automatic tree detection algorithm, it was possible to create baseline tree density layers for the 2 years of observation and to quantify the tree cutting rates for the period from 2001 to 2006. An average annual tree loss of −2.8%, coupled with the total absence of regrowth during the 5 years study period, confirm the tremendous ecological impacts of charcoal driven tree cutting on tiger bush vegetation. Analysis of the results evidences spatial and temporal patterns in the cutting locations and poses the basis for a better understanding of the ecological and human dimensions of deforestation in the fragile rangeland environment of Northern Somalia.

The influence of conifer forest canopy cover on the accuracy of two individual tree measurement algorithms using lidar data

Individual tree detection algorithms can provide accurate measurements of individual tree locations, crown diameters (from aerial photography and light detection and ranging (lidar) data), and tree heights (from lidar data). However, to be useful for forest management goals relating to timber harvest, carbon accounting, and ecological processes, there is a need to assess the performance of these image-based tree detection algorithms across a full range of canopy structure conditions. We evaluated the performance of two fundamentally different automated tree detection and measurement algorithms (spatial wavelet analysis (SWA) and variable window filters (VWF)) across a full range of canopy conditions in a mixed-species, structurally diverse conifer forest in northern Idaho, USA. Each algorithm performed well in low canopy cover conditions (<50% canopy cover), detecting over 80% of all trees with measurements, and producing tree height and crown diameter estimates that are well correlated with field measurements. However, increasing tree canopy cover significantly decreased the accuracy of both SWA and VWF tree measurements. Neither SWA or VWF produced tree measurements within 25% of field-based measurements in high canopy cover (i.e., canopy cover >50%) conditions. The results presented herein suggest that future algorithm development is required to improve individual tree detection in structurally complex forests. Furthermore, tree detection algorithms such as SWA and VWF may produce more accurate results when used in conjunction with higher density lidar data.

Using Laser Altimetry-based Segmentation to Refine Automated Tree Identification in Managed Forests of the Black Hills, South Dakota

The success of a local maximum (LM) tree detection algorithm for detecting individual trees from lidar data depends on stand conditions that are often highly variable. A laser height variance and percent canopy cover (PCC) classification is used to segment the landscape by stand condition prior to stem detection. We test the performance of the LM algorithm using canopy height model (CHM) smoothing decisions and crown width estimation for each stand condition ranging from open savannah to multi-strata stands. Results show that CHM smoothing improves stem predictions for low density stands and no CHM smoothing better detects stems in dense even-aged stands, specifically dominant and co-dominant trees (R 2 � 0.61, RMSE � 20.91 stems with smoothing; R 2 � 0.85, RMSE � 46.02 stems with no-smoothing; combined smoothed CHM for low density and unsmoothed CHM for high density stands R 2 � 0.88, RMSE � 28.59 stems). At a threshold of approximately 2,200 stems ha � 1 , stem detection accuracy is no longer obtainable in any