Recipient Organization
UNIVERSITY OF MAINE
(N/A)
ORONO,ME 04469
Performing Department
School of Forest Resources
Non Technical Summary
Since 2000, a large body of remote sensing (RS) sensors and platforms has become available to overcome shortcomings related to the unavailability of optical RS imagery and their cost. These sensors can produce enhanced geospatial data products to meet research and decision support needs. Despite these considerable advances, RS methods to adequately apply these data by combining several sensors and platforms are not yet developed in particular for forestry applications. Consequently, the long-term goal of this project is to develop sound approaches for providing detailed geospatial products on forest tree identification/composition and forest defoliation/damage caused by recent destructive pest/pathogen outbreaks. Rather than a single use technology, the continually updated products offered here will both support a strong and sustainable research program at the University of Maine (UM) as well as provide stakeholders with enhanced forest health information needed for decision making.
Animal Health Component
70%
Research Effort Categories
Basic
30%
Applied
70%
Developmental
(N/A)
Goals / Objectives
Main goal: Develop a comprehensive research framework for forest health/condition assessment and to provide products for decision making. Specific objectives are:1) Select and design several experimental sites in Maine to collect ground-level data to support future remote sensing research requirements and select detectable vegetation biophysical and biochemical parameters2) Develop tools using remote sensing spectral information and machine learning algorithms to detect and map defoliation and damage related to current pest/disease outbreaks in Maine3) Develop methods to map forest composition and tree species for Maine using fusion of several forms of remote sensing data4) Establish and implement a mechanism to share data and products with the Maine Forest Service and forest resource stakeholders and receive feedback to improve the process for development of tools and methods
Project Methods
Our research methods to address our objectives are composed of the following components:1. Developing models for tree-level species mapping:We hypothesize that individual tree species can be mapped more effectively using a combination of remote sesning (RS) spectral information and other environmental data such as slope, elevation, depth to water table, and tree phenology to result in more reliable predictions.A similar approach can be applied for finer resolution, multispectral data such as WorldView-3 or hyperspectral imagery to map species such as black ash vs. other ash species and eastern white pine. In this study, a new method will be developed to map individual trees using multisource data including remotely sensedspectral data, site-level characteristics, and tree phonological information. Both G-LiHT hyperspectral imagery and WorldView-3 (http://worldview3.digitalglobe.com/) will be tested and compared. The performance of the pixel-based and object-based approacheswill be separately evaluated in our study sites. WorldView-3 and 4 images will be purchased for several sites that will be identified based on data presented by Costanza et al (2015) and from University of Maine experimental forests such as the Penobscot and Howland locations. Already available G-LiHT imagery will be used and new G-LiHT data will be collected for the selected sites if not already available. Advancement of tree species identification is critical in Maine for the number of species present, the complexity of the forest as a result of past/current management, and vast array of disturbance agents on the landscape. The developed models will be applicable to other regions with similar forest attributes.Field data on some of the tree species proposed to be used for model training and accuracy assessment are already collected by the Maine Forest Service (MFS) or available from other sources.However as explained in the objectives of research section, new study sites will be established and additional data will be collected. Although the focus is on model development, the final products of this component will be produced for areas recommended by collaborators to be shared with stakeholders such as Emerald Ash Borer (EAB) Task Force for initial planning.2. Developing models for crown-level ash infestation and white pine decline:Pest- or pathogen-induced stress in ash or eastern white pine trees can be detected before defoliation is visible by estimating vegetation stress through changes in a plant's physiological processes such as photosynthesis and/or evapotranspiration rate using chlorophyll fluorescence and thermal sensors, respectively. Stress can cause an increase in leaf and canopy temperature and can often be detected by thermal infrared imaging as the early stage of infection develops. There are models based on using thermal infrared data alone or in combination with VIs for vegetation water stress detection. Ishimura et al. (2010) applied the model developed by the PI to detect beech forest damage due to ozone pollution using MODIS imagery. Although thermal infrared data have been applied for vegetation disease and stress detection in agriculture at leaf and local scales, research is limited for forestry applications. Smigaj et al. (2015) applied UAV-based thermal infrared data to detect increase in temperature related to the disease progression in Scots pine (Pinus sylvestris L.) due to red band needle blight (Dothistroma septosporum) infection. The G-LiHT platform will soon add a solar-induced chlorophyll fluorescence (SIF) sensor that will provide a unique opportunity to acquire information through all four sensors for forest health assessment.Our hypothesis is that tree level infestation can be detected more effectively using our suggested methods however this research will also evaluate the effectiveness of the methods to accurately separate infestation from other drivers of damage. Two different approaches will be evaluated in this project for tree-level infestation detection in ash: 1) G-LiHT hyperspectral, thermal infrared, and SIF data will be assessed to detect early infestation (before defoliation is visible) and 2) defoliation and mortality will be estimated using VIs obtained from more representative Worldview bands. The field data for training and accuracy assessment will be collected over the first two years of the project in areas in Maine and neighboring states where infestation has been observed by collaborators.3. Landscape-scale detection and severity classification of annual SBW defoliation: A method has recently been developed by the PI to detect and quantify the current-year (annual) SBW defoliation using Landsat imagery (30 m spatial resolution) using SBW outbreak data from Quebec where the outbreak is coming from. The PI and her team synthesized several data sources including RS imagery to model current-year SBW defoliation. Seven VIs were tested for their capacity to detect and quantify defoliation using the random forest (RF) non-parametric regression method. Our results indicate that the combination of NDVI, enhanced vegetation index (EVI), and normalized difference moisture index (NDMI) can reduce the OOB (out-of-bag) rate of error for defoliation detection by 5% and classification by 9%. Building on our previous study, the following outlined research needs to be implemented to produce annual SBW defoliation data for operational and scientific use with higher probability of productionand lower commission error.The current model based on Landsat-5 data is capable of classifying defoliation severity into three classes (light, moderate, and severe). However, both Landsat-8 and Sentinel-2 data have higher quality (signal to noise ratio) and radiometric quantization compared to previous Landsat instruments. It is expected that defoliation severity can be classified finer than in past efforts (e.g. 4-5 classes). This need has been frequently emphasized by stakeholders, and current RS or aerial surveys are unable to address the information gap. In this project we will train the model using pixel-based data using a tree health assessment method to be developed that uses a combination of field data and high resolution hyperspectral/multispectral aerial imaging acquired by University of Maine using a similar method developed by Brovkina et al. (2017). The Barbara Wheatland Geospatial Analysis Program (BWGP) in SFR operates a high-resolution multispectral camera onboard a Cessna aircraft and a UAV system which will be used for data collection at single tree level during the first two summers of the project. Acquired aerial imagery will be pre-processed by BWGP before final applications. Field visits will be conducted to support the development and validation of the suggested tool in collaboration with MFS experts. Several locations have been already identified as SBW hot spots in northern Maine for the project's field work.Two approaches are suggested to reduce the error of commission: 1) ancillary data such as RS-derived tree species map that will be produced in this project and 2) classifiers such as multi-criteria support vector machines will be evaluated (MCSVM). The final step is to automate all the processes explained above to develop a rapid, cost-effective, and highly accurate tool to produce data on annual SBW defoliation for the state of Maine for the first three years of the outbreak.