Recipient Organization
PURDUE UNIVERSITY
(N/A)
WEST LAFAYETTE,IN 47907
Performing Department
Forestry & Natural Resources
Non Technical Summary
Forests provide critical ecosystem services but are impacted by intensified environmental challenges. Precision forest management and tree improvement efforts both require precise and detailed information. However, management decisions are often made using incomplete or qualitative data, hindering the ability to make truly informed, data-driven decisions that improve forest environmental and economic sustainability. Many conventional methods are not feasible over large areas (Hilkers et al. 2012), require considerable manual effort, thus incurring high costs (Estornell et al 2015, Xiao et al. 2005). In addition, they often rely on observations by trained experts, necessitating on-site inspection of individual trees, introducing substantial sources of error and reducing reproducibility of data collection efforts. Meanwhile, significant improvements have been made in digital technology and remote sensing that helped to lay the foundation for precision management. However, a critical technological gap remains in capturing key data and integrating multi-resolution, multi-stream and multi-platform data that can produce locally operationable information for precision management. The project will enable the use of more sophisticated data acquisition tools among the broader professional and public communities for precision forest management that will be applicable in most temperate forests in the U.S. and beyond.
Animal Health Component
30%
Research Effort Categories
Basic
40%
Applied
30%
Developmental
30%
Goals / Objectives
The overarching goal of this proposed project is to develop and test a set of digital tools to modernize forest inventory and management. The specific objectives for this project are:Obj. 1. Develop an AI-assisted application for tree species identification based on bark (Co-PD Gazo). We will develop an annotated bark image bank of hardwood trees in the Central Hardwood region, an AI algorithm for bark-based species recognition, and a smartphone app for bark-based tree species identification. The smartphone app will be freely available to the public and AI algorithm can be applied in precision forest management.Obj. 2. Develop tools for automated detection and delineation of individual trees for hardwood species using low-density aerial LiDAR (Co-PD Hardiman and Shao). We will use statewide airborne LiDAR data captured in 2011- 2013 and in 2016-2018 to develop and improve two types of algorithms, canopy-height model and point-density model, for the delineation of individual trees. Tools developed from this objective can be applied at stand, landscape, and possibly state level using freely available aerial LiDAR. Obj. 3. Develop tools for automated detection and delineation of individual trees and measurement of biometrics for hardwood species using high-density terrestrial LiDAR (Collaborator Habib). We will develop a new terrestrial high-density LiDAR system to conduct effective under canopy application. And we will develop new algorithms for individual tree identification, georeferencing, and parameterization. Tools developed from this objective can be applied at plot, stand, and landscape level for precision management. Obj. 4. Develop algorithms for automated detection and delineation of individual trees and measurement of biometrics for hardwood species using UAS orthophotos (Collaborator Hupy and Jung). Visible-band sensors will be used for the airborne data acquisition. 3D digital surface models (DSM) will then be created based on the orthophoto mosaic, and image fusions of the orthophoto mosaic and DSM will be conducted to facilitate tree identification, delineation, and biometrics estimation. Tools developed from this objective can be applied on the stand level and can be employed cheaply and as frequently as the user desires.
Project Methods
Objective 1. AI-assisted application for tree species identification (Gazo, Fei & Shao)Three key components will be completed for species ID: an annotated bark image bank of hardwood trees, an AI algorithm for bark-based species recognition, and a smartphone app.Data Acquisition A Bark Image Collection App (BICA) for smartphones will be developed. The app will prompt user to take picture of a tree from a certain distance, at DBH height and from 4 perpendicular directions. At the same time, GPS location, camera type, focal length, aperture, ISO, compass heading and other metadata such as environmental and ambient factors will be recorded for each image. A prototype of the image collection App has been developed.Image bank and annotation The dataset will be gathered on a Purdue server and curated. All images will be manually verified. The curated dataset will be prepared for deep learning (with and without background clutter). We will use Res-NET to segment the tree from the background. Then we will use Convolutional Neural Network (CNN, see paragraph below) to classify the bark into the tree category. We will parse the images and subdivide into patches of 256x256 that are suitable for deep learning.Tree identification algorithm and application Trees species can be identified in several ways. Our focus is on using tree bark for identification purposes. Bark images are considered a texture - systematic local variation of image values (Remeš et al. 2019). In computer vision and image processing fields, texture analysis is a basic topic. All image texture analysis methods can be generally divided into these categories: 1) statistic-based methods, 2) model-based methods, 3) signal processing-based methods and 4) structure (math) based methods. CNN, the currently most-used approach in texture analysis, can be considered a model-based method. CNN is a deep learning architecture inspired by the natural visual perception mechanism of the living creatures. Recent CNNs are comprised of groups of convolutional, pooling, activation, and fully-connected linear functions and they include hundreds of thousands connections (Svab 2014). A smartphone app will be developed and made freely available to the public for use in precision forest management.Objective 2. Low-density aerial LiDAR tree delineation and measurementLiDAR provides digital height models that can be used to identify and measure tree biometrics at a broader geographic scale, which are essential for to estimate key ecosystem services such as carbon pool sizes.LiDAR Data We will use statewide airborne LiDAR data captured in 2011- 2013 and in 2016-2018. The 2011-2013 data were collected with Leica ALS50 (Leica Geosystems) during the leaf-off season. This laser sensor collected up to 4 returns per pulse with 99 KHz of pulse repetition rate, flying at the above ground level (AGL) of 2,000 m with the scan angle of 40 degrees and scan frequency of 35.8 Hz. The average density of the point cloud in the study sites was approximately 2.3 pts/m2. The 2016-2018 data, which are partially available currently, have an approximate point density of 8 pts/m2.Individual tree identification Two groups of algorithms are available to delineate individual trees. Canopy-height-based (CHB) algorithms work well in coniferous forests or sparse deciduous forests with easily discernable crown tips (Kaartinen et al. 2012). Point-density-based (PDB) algorithms may overcome the deliquescent tree forms in deciduous forest but have only been tested with high-density LiDAR data (Rahman and Gorte 2009). We propose a 2D point-density-based algorithm to identify individual trees with low-density LiDAR. We will apply a marker-controlled watershed segmentation strategy to delineate individual crowns with PDB models (Chen et al. 2006). We propose to expand upon this preliminary study to automate the tree identification algorithm and scale up for application in various HTIRC plantations. We will also test the feasibility of the algorithms developed in this project in natural forest plots that are monitored by the IN DNR Continuous Forest Inventory (CFI) program.Tree attributes extraction Tree-level attributes, tree height, crown width and volume will be extracted and estimated after individual crown delineation. Tree height information can be extracted by LiDAR point height attributes. Crown width will be calculated by using the area of the crown segments. We will then collect DBH on the ground to build the crown width-DBH relationships, which should be feasible based on our preliminary study. We will first perform these analyses on hardwood tree plantations in our research forest, Martell forest, and then expand to natural forests.Objective 3. High-density terrestrial LiDAR tree delineation and measurementHigh-density LiDAR can provide biometrics measures that are previously not possible such as tree allometry and various traits analysis. We will develop a customized hardware system to capture and related software to extract key features.System integration, data logging, and deployment: Prototype systems have been recently integrated by our team. We will finalize the integration of these platforms (including the synchronization among the different system components - GPS, Camera, and LiDAR; data logging; and system calibration to precisely estimate the mounting parameters relating these components). In parallel, the systems will be deployed in Martell Forest for the collection of image and LiDAR data and fine tuning the system integration parameters according to preliminary analysis of the collected data.Develop data processing and reduction strategies: Acquired data by the camera, LiDAR, and GPS units will undergo a system-driven approach for data process, where we will develop an optimization procedure that integrates multi-modal remote sensing data with GPS position and orientation information for the derivation of precisely georeferenced. This process will also involve the optimization of the GPS trajectory using image and LiDAR data to mitigate the impact of GNSS-signal blockage by canopy. In addition, data reduction strategies will be developed for the identification of individual trees and derivation of their biometrics. We will also compare data collected by the proposed systems with other data sources (e.g., static LiDAR, UAV-based LiDAR, statewide LiDAR from manned aircrafts, and high-resolution satellite imagery). Visualization tools will be also developed for interactive manipulation of multi-modal/multi-platform/multi-temporal geospatial data/products.Objective 4. UAS orthophoto tree delineation and measurementImage acquisition A visible-band sensor will be used for the airborne data acquisition, which will be mounted on a DJI Phantom 3 Advanced multirotor aircraft the team has. The reason we chose the specified platform is because of its affordability, compatibility with autopilot software, and capable 12 megapixel camera, which also likely can be afforded by landowners and other insitutions. We will use DroneDeploy to program the flight and trigger the onboard camera.Image processing Individual images will be combined to form a true orthophoto mosaic of the forest surveyed at a spatial resolution of 2.5 cm per pixel. A 3D digital surface model (DSM) will then be created based on the orthophoto mosaic using photogrammetry software Pix4D. An image fusion of the orthophoto mosaic and DSM will be conducted in Erdas Imagine software. The whole process can be done in readily available programs and the feasibility has been tested in our pilot study.Individual tree identification and attributes extraction In this project, we will develop new algorithms to automate the process. Given the 3D DSM has the similar attributes to the LiDAR based digital crown model, but with much higher point density (160 pts/m2), we intend to test both the canopy-height-based algorithms and Point-density-based algorithms.