Recipient Organization
MICHIGAN TECHNOLOGICAL UNIV
1400 Townsend Drive
HOUGHTON,MI 49931
Performing Department
College of Forest Resources and Environmental Science
Non Technical Summary
Individual tree detection and delineation (ITDD), as the name suggested, is the technique that aims to locate and delineate individual tree canopies. Reliable ITDD technique would not only tell the total number of trees in a forest, but also provide the dimension of canopy for each detected tree. These data products would pave the way for precision forestry management(Gougeon and Leckie 2001; Holopainen et al. 2014), increasing the profitability of forestry by allowing better allocation and optimization of various forestry-management tasks and loggings.Due to the difficulty of obtaining individual tree canopy measurement in traditional field inventory, most models that study forest structure and dynamic still depend on diameters at breast height (DBH). However, recent studies have shown that including canopy size into the tree biomass estimation model would substantially improve the accuracy of tropical estimates of tree biomass in primary and degraded forests(Goodman et al. 2014). Besides, an increased interest of individual tree canopy size has appeared in other studies of forest in the past decades(Bohlman and Pacala 2012). ITDD is an important step for tree species classification, which turned out to provide important information for accurate estimation of tree carbon stock(Goodman et al. 2012). Previous studies also indicate that allometric relationship of canopy size and DBH shows no significant inter-site variation(Blanchard et al. 2016), meaning that the ITDD product could also be used to derive reliable DBH information.With the immense interests of obtaining individual tree crown information from both forest managers and forest ecology researchers as discussed above, ITDD using remote sensing data is still an unsolved problem. Forest inventory based on single tree information have failed to challenge the area-based approach (ABA), primarily because reliable ITDD for the detection of individual tree in various forest conditions does not exist (Kaartinen et al. 2012).ITDD often relies on LiDAR derived raster data called canopy height model (CHM) to detect the tree tops, based on which tree crown delineation is conducted with various approaches including region growing, energy balance, watershed, hill climbing and valley following. Tree top detection is the key for achieving good ITDD results. Tree top detection is usually performed by local maxima extraction procedure. While this method works well for coniferous forest, it tends to fail for deciduous forest, since deciduous trees do not satisfy the assumption that one tree only has single local maxima: deciduous tree usually has multiple tree tops detected using local maxima method. In an attempt to solve this problem, (Liu et al. 2015) proposed to train a classifier to categorize the boundaries between two neighboring tree tops into two types: boundary between trees and boundary within a tree. If the boundary is classified as the boundary within a tree, then the tree tops on the two sides of the boundary will be merged to form one tree top. After tree top detection and merging is finished, an algorithm called Fishing Net Dragging (FiND) was developed to delineate the tree canopy. While this ITDD approach seem promising with better performance than existing methods, collecting reliable training samples to train a robust boundary classifier is challenging for practical use. Some other researchers tried to improve the accuracy of tree top detection by removing the false tree tops using tree allometry rules. The performance of this method is relying the accurate estimate of several parameters that control the rules. However, those parameters require the users' input and are usually site specific, making it difficult to obtain good performance when adopted in practice(Sa?kov et al. 2017). One of latest implementations of ITDD using CHM is found in (Yin and Wang 2019), which shows the difficulty of obtaining reliable ITDD results in dense forest from airborn LiDAR data and suggests including terrestrial LiDAR data and spectral remote sensing images into the ITDD procedure as future study.Another technical direction of ITDD is to segment the raw LiDAR point cloud directly into individual tree canopies using segmentation method like mean-shift algorithm(Ferraz et al. 2016) and graph-based segmentation algorithm(Strîmbu and Strîmbu 2015). This type of method is named point-based method, in contrast with the ones that are based on rasterized CHM as discussed above. Two benchmarking studies conclude that the point-based methods have better performance than the CHM raster-based methods and suggest future research and development efforts need to be invested in the point-based methods(Wang et al. 2016) and point cloud and raster data be combined for potential further improvements(Xiao et al. 2019).An emerging direction in the ITDD research area is to utilize the deep learning techniques. (Weinstein et al. 2019a) proposed a semi-supervised method to detect individual tree crown. This method can be summarized the three steps: 1) initial tree location generation using unsupervised method, 2) train an object detection model (deep learning model) with the noisy samples generated from step 1, 3) fine-tuning the model using hand-annotated samples. While this method seems promising as it indicates it can be applied to a wide array of forest types at broad scales with the experiment results for sites of National Ecological Observatory Network(Weinstein et al. 2019b), there still some limitations of this approach. First, the method adopted by this ITDD approach to generate initial tree locations relies on the rasterized CHM, while many studies indicate that point-based method tend to provide more accurate tree locations. Second, only RGB images are used in the deep learning model training procedure, but the point cloud that characterizes the internal canopy structure is not utilized in the deep learning model. Therefore, this ITDD tends to fail to detect small trees under the big canopy as shown by the author(Weinstein et al. 2019b). In addition, hyperspectral or multispectral data has not been used in the model training, even though the abundant spectral information is expected to help differentiate neighboring canopies.Justification of proposal: To date, there is no operational and easy-to-use method available for ITDD, even though there is a high demand for ITDD product from forest owners and managers, and researchers that focus on the trees in natural or urban environment.With the latest deep learning techniques and point-based tree detection method, and the increased availability of LiDAR and hyperspectral data, this project proposes new method to solve ITDD problem, with the ultimate goal of developing a model that has a user-friendly interface and can provide reliable ITCD in various types of forest.List of reference will be provided in the attached proposal document
Animal Health Component
20%
Research Effort Categories
Basic
0%
Applied
20%
Developmental
80%
Goals / Objectives
1.Develop a novel individual tree detection and delineation (ITDD) for Ford Research Forest.2.Develop a universal ITDD model for National Ecological Observatory Network3.Use ITDD product to find better solutions to problems in ecology and silviculture.
Project Methods
Year 1 (December 2020 - November 2021) The current FRF dataset will be reviewed to verify if they satisfy the research needs. Otherwise, new dataset or field work need to be conducted. Assuming the qualified dataset is in place, the research activities will include 1) testing different point-based and raster-based algorithm to generate initial tree locations, 2) finding the best algorithm from step 1 to generate the initial tree locations, 3) investigate different ways to fuse LiDAR, UAV and hyperspectral data to create training samples, 4) customize a deep learning model to make it able to take those sample 5) train the deep learning model using those samples, 6) investigate how finetuning of the model using hand-labelled samples impacts the model performance.Year 2 (December 2021 - November 2022) NEON provides RGB, LiDAR and Hyperspectral data for its sites, which should work with model developed in the first year. The primary activities in year 2 include 1) download the NEON data to the local machine, 2) collect samples from NEON for training and evaluating the model 3) apply the trained model to all the sites of NEON 4) evaluate the performance of the model for NEON data 5) contact the NEON to publish the ITDD product on NEON website for free public use to generate the impact of our work.Year 3 (December 2022 - November 2023) The focus of year 3 is to work with researchers with ecology and silviculture backgrounds to identify the problems in their research areas that can incorporate the ITDD product and investigate if a better solution can be obtained to solve those problems after adding ITDD product to their solutions.