Recipient Organization
UNIV OF IDAHO
875 PERIMETER DRIVE
MOSCOW,ID 83844-9803
Performing Department
FRFS
Non Technical Summary
Revisioning Precision Forestry designs, develops, and tests the accuracy and precision of novel imaging technologies using drones (UAS) and other systems to implement improved processes for derivingforest inventory metrics within operational management standards. Our priority is to resolve constraints on access and use of terrestrial and aerial remote sensing technology to monitor, measure, and manage western United States forested ecosystems, which are under significant stress due to climate change, wildfires, insect attacks, increasing wood utilization demand, and a burgeoning carbon offset market. We propose to reduce access constraints by developing and publishing automated, open-access software algorithms forprocessing terrestrial and UAS remote sensing data into formats that aid visualizing, managing, and communicating forest stand and tree attributes using currently available, open-source forest modeling software. Specifically, we envision democratizing precision forestry through exploring novel principles, such as allometric relationships, to more precisely fuse (co-register) above and below canopy point clouds derived from terrestrial and UAS remote sensing platforms. We will then assess the automation of deriving individual tree and forest stand attributes from fused point clouds that would be directly ingestible by the Forest Vegetation Simulator program, the most widely used forest growth and yield software program within the United States. Lastly, we will evaluate the cost/benefit of utilizing photogrammetry alone to estimate forest and tree attributes, relative to more costly lidar remote sensing platforms. We will use advancements in computational machine learning (e.g., XGBoost), coupled with open-source analytical platforms (SQL, Python, R) to test and develop new methods for modeling the forested environment through readily accessible and relatively inexpensive remote sensing technologies.This research project is designed to solve a key national problem limiting the optimization of forest resource management. Namely, that many organizations and land managers have limited or no access to software support systems that enable comprehensive remote sensing quantification of tree and forest attributes. Accurate and comprehensive quantification of forest and tree attributes, in actionable and meaningful formats, is critical to proactively maintaining healthy and productive forests, capable of supplying long-term ecosystem services (e.g., fiber, carbon sequestration, habitat, water quality, cultural traditions). We intend to overcome these software and data access limitations by developing open-access software algorithms for processing point clouds of forested environments. Open-access is critical, as the threats and opportunities facing our forested ecosystems and the services they provide culturally, ecologically, and economically are only increasing. Access should not be limited to those with significant financial resources, but to all who have an interest and responsibility to manage our national forest resource. Project outcomes will provide software tools accessible by all stakeholders in the forest management community to allow precision forestry to 1) Promote sustainable forest ecosystem services, 2) Sustain natural resource-based economies, and 3) Advance the integration of technology into organizational forest management workflows.
Animal Health Component
100%
Research Effort Categories
Basic
(N/A)
Applied
100%
Developmental
(N/A)
Goals / Objectives
Ourgoals are to: 1) democratize precision forestry through testing and validating low-cost off the shelf remote sensing platforms to accurately measure forest attributes above and below forest canopy (stem size, species, and location); 2) develop open-source algorithms for co-registering above and below canopy point clouds derived from lidar and photogrammetry; and 3) create an open-source software to convert point clouds into tree lists digestible by forest modeling platforms. These goals will be achieved through the completion of the following four objectives.Objective 1 seeks to enhance existing point cloud co-registration algorithms by integrating allometric models into the alignment process. We hypothesize that the allometric informed approaches will reduce above and below canopy point cloud co-registering errors. We will deliver open-source software allowing any individual or organization to produce an allometric informed co-registered point cloud collected above and below a forest canopy.Objective 2 will assess tree measurement accuracy from co-registered above and below canopy point clouds derived from Structure from Motion (SfM - inexpensive) against co-registered below canopy handheld mobile lidar scanning (HMLS - expensive) and above canopy SfM. Accuracy comparisons will occur across multiple forest productivity zones and species compositions in the western U.S. Stem mapped plots will be used to evaluate platform accuracy across Idaho, Oregon, and Colorado within Douglas-fir and ponderosa pine forests - the two most ecologically and economically important species in the western U.S.Objective 3 will perform a cost-benefit analysis of the remote sensing platform accuracy, error propagation, acquisition time, and tradeoffs between equipment cost and post-acquisition processing requirements.Objective 4 will develop an open-source software package to process SfM/HMLS derived point clouds, co-register, and produce an individual tree list for the scanned area. This tree list will be formatted to be compatible with the most widely used forest modeling software package in the U.S. - the US Forest Service Forest Vegetation Simulator.
Project Methods
Methods - Objective 1:The proposed methods will operate iteratively through the following four steps: 1) Generate local predictions of DBH using the heights of each tree extracted from the UAS-SfM dataset, where the prediction will be based on data from freely availableU.S. Forest Service Forest Inventory and Analysis data, a dataset containing more than 100,000 forest plots across the U.S. 2) Match the predicted DBH values for the UAS-SfM dataset with DBH values extracted from the terrestrial point cloud so that they minimize values extracted from the terrestrial point cloud and minimize DBH error. 3) Determine the x,y,z transformation that will minimize the distance between the matched predicted and extracted DBH values. 4) The resulting transformation will then be applied to the point clouds to generate a single co-registered dataset using the Maximum Likelihood Estimation Sample Consensus (MLESAC) algorithm, which will be classed into ground and tree architecture returns. We will assess the effectiveness of the proposed co-registration algorithm by assessing tree location accuracy within the individual above and below canopy datasets against the accuracy within the co-registered dataset. These accuracies will be quantified by comparing remotely sensed trees to the stem mapped trees. Although we will assess a suite of different RMSE metrics, vertical error metrics will include deviations from the tree base and tree height elevations, and horizontal error metrics will include deviations from the centered location of each tree.Methods - Objective 2:In each of the 16 study stands, traditional fixed area plots will be installed and located using a RTK GPS to ensure the data can be spatially aligned with remote sensing observations for validation. At each fixed plot area, all trees will be spatially located (elevation of tree base and x,y coordinate) and have their DBH, upper stem diameter (5.18 m above ground), height, crown base height, and species recorded. Shrub cover and height will be recorded and locations of large, downed wood to assess co-registration variability. These observations will allow for thevalidation of the remotely sensed tree-level parameters, such as DBH, height, and timber volume, but will also provide traditional stand-level summary estimates of density (i.e., basal area and trees per hectare) and biomass (i.e., cubic meters and tons per hectare of timber) for comparison with remote sensing estimates.Each of the study stands will be photographed using a Sony Alpha 7II with a 14 mm lens connected to an Emlid M2 RTK transponder to geotag the images. Photos will be collected while walking parallel paths through the stand spaced approximately 10 m apart. These photos will be processed using Agisoft Metashape Structure for Motion algorithm to generate high density terrestrial point clouds. Through segmentation of individual trees within the point cloud, characteristics, including DBH and upper stem diameter, will be extracted. A Zeb Horizon HMLS capable of measuring up to 300,000 points per second in single return mode within a maximum range of 100 m will be used to scan each of the 16 study sites. Scanner paths will be spaced approximately 20 m apart to ensure optimal coverage of each study area, obtain low scanner range noise, and reduce the drift associated with the receiver algorithm. The HMLS field data will be processed in GeoSLAM Hub software using the SLAM algorithm to integrate IMU positional data and calculated trajectory with distance and angle of each point to determine theirrelative position. Resulting point clouds will then undergo identical processing as the T-SfM data for extracting tree stem metrics.Each study stand will be flown with a DJI Matrice 210 RTK multirotor UAS equipped with a Zenmuse X5S camera (20.8 MP, 15 mm lens). Images will be acquired in a serpentine pattern with 90% forward and 85% side image overlap. The UAS will be operated at ~90 m altitude and a flight speed of 4 m s-1. Acquired imagery will be processed using Agisoft Metashape to generate high density point clouds for maximizing tree extraction from the resulting canopy height model (CHM). The site CHMs will be processed using a variable window function (VWF) to identify individual tree locations and heights.After data collection, the UAS-SfM datasets will be co-registered in separate steps with the coincident T-SfM and HMLS datasets to create two datasets for each site containing the above and below canopy point clouds. Extracted tree locations will be compared against field stem mapped tree locations to quantify omission and commission of individual trees, which will be analyzed against tree size and local forest density metrics to understand how forest structure impacts the success of remote sensing tree extraction. Additionally, individual tree DBH, upper stem diameter, taper, height, and volume from the remotely sensed datasets will be compared to the stem mapped values to quantify tree-level accuracy. We will compare remote sensing and stem mapped parameters using regression-based statistical equivalence testing. Further, we will test for intercept and slope equality between the sets of measures of the same parameter.We will assess individual species through constraining RandomForest (RF) classification predictions to potential species that are likely present within a stand due to soil type, latitude, elevation, aspect, moisture availability, etc. We will integrate forest attribute data, point cloud-derived structural data, and spectral datasets from the UAS systems. Through the RF approach, we will generate 1000 classification trees from a random 60% subset an calculate bootstrap error estimates by classifying 40% of the remaining training data. Based on the bootstrap calculations, we will determine species specific True Positive and False Positive rates.Methods - Objective 3:We will conduct a cost-benefit assessment of pure SfM and SfM+HMLS remote sensing technologies against traditional forest inventory strategies.During remote sensing data collection and processing, detailed records of vested time will be tracked and standardized as a measure of hours per acre so that direct comparisons of invested time can be made. These comparisons will also be compared against published standards for traditional forest inventory methodologies. Since the largest component of these remote sensing systems is expected to be the cost of terrestrial data collection, we will augment this effort by investigating how the intensity (e.g., proportion of area sampled) and strategy (i.e., strip, circular, square plots) of terrestrial remote sensing data collection impacts the cost-benefit relationship of precision forestry from remote sensing. This will be done by collecting 100% census of terrestrial data at each of the study sites and then resampling these data to represent different collection strategies. This will enable us to determine what percentage of the area needs to be sampled to still attain metrics to a precision acceptable to our stakeholders, thereby assessing realistic estimates of future cost. Cost will be measured as both fixed (i.e., equipment) and variable (i.e., collection and processing time) against different derived metrics of forest inventory precision.Methods - Objective 4:A simple graphical user interface (GUI) will be developed, which wraps the co-registration (Objective 1) and tree attribute derivation algorithms (Objective 2) into a package that allows the user to input co-registered point clouds, specify parameters, visualize the segmented trees, check for accuracies, and output tree attributes to an FVS-ready comma separated value (.csv) format. This GUI will allow the user to have control of the process using buttons, drop downs, checkboxes, and other control options typically found in interface design.