Progress 10/01/19 to 09/30/21
Outputs
Target Audience:The primary target audience was forestry and other natural resources practitioners. Secondary audiences were Researchers working in forest ecosystems. Changes/Problems:As described in other sections of this report, the mobile laser scanning approach was genererally unsuccessful because we were never able to consistently obtain good point clouds from the equipment. We tried multiple SLAM algorithms, and none of them consistently converged, and when they did converge, they included "phantom trees" that resulted in false tree identifications (false positives). Part of the problem is with the SLAM algorithms, which work best in environments with large, regular features, such as buildings and walls. They do not work as well in the relatively chaotic environment of a forest. These limitations may have been overcome with a higher quality inertial measurement unit (IMU), but we did not have the resources to purchase that hardware. Furthermore, lower-cost commercial mobile laser systems are coming on the market now, so further development on our own system did not seem to be the best use of our resources. As a result, we shifted our focus to using the available ground data and high-quality airborne LiDARdata to develop a new individual tree identification algorithm. What opportunities for training and professional development has the project provided?One master's student was trained as a result of this project. Also, the research technician gained considerable technical skills as part of the project. How have the results been disseminated to communities of interest?We have one research paper (by J. Hershey et al) in review. J. Hershey also gave a presentation at thePennsylvania Forestry Association's 2021 Annual Symposium entitledLiDAR-based Individual Tree Detection for Forest Inventory and Carbon Quantification. We have a proposal in with the DCNR Bureau of Forestry to use airborne and mobile LiDAR to enhance their forest inventory proceedures. What do you plan to do during the next reporting period to accomplish the goals?
Nothing Reported
Impacts
What was accomplished under these goals?
1. Develop algorithms to process LiDAR point clouds in order to obtain tree metrics. Remote sensing research oncharacterizing forest ecosystems and associated metrics hasresulted in a varietyofopen-source processing tools. However, the developersof these tools havegenerally focused on homogenous even-aged stands, and their application in more complex forests, such as Eastern hardwoods,leave much to be desired. In addition, functions are spread amongst a variety of software libraries and are generally disjointed due the complex nature of processing 3D point cloud and 2D spectral data. Our teamfocused on distilling the breadth of available tools and libraries into succinct processing pipelines. We used these tools to develop an agorithmto 1) construct point clouds from the raw data obtained fromour backpack LiDAR system, 2) identify and 3) measure tree stems within the resulting point clouds. Point clouds were constructed from the raw data usingSimultaneous Localization and Mapping(SLAM)algorightms.Stem identification was based on the Hough transform. Point cloud data were first voxelized, and then voting criteria based on local point density maxima and vertical continuity were used to identify stems. A random sample consensus (RANSAC) cylinder fitting algorithm was used to estimate tree stem diameters. 2. Assess the repeatability,accuracy and precision of tree metrics obtained from LiDAR point clouds. The hardware and algorithms developed in this study were used to collect LiDAR data, create point clouds,and identify and measure trees on 6425x25m plots. Prior to scanning, we mapped and measured each tree on the plots using conventional ground-based techniques. Trees were located with an average precision of ~11cm, and DBHs were measured to the nearest cm.We scanned the plots with the mobile laser system multiple times, but we were not always able to get the SLAM algorithms to converge to produce a coherent point cloud. Acceptable point clouds were only obtainedin a small percentage (<20%) of the cases. In most cases when usable point clouds were obtained,100% of trees >4" were identified by the algorithm. However, the algorithm also tended to identify"phantom trees," i.e., false positives. In the best cases, as few as5% of trees identified werefalse positives, while in other cases as many as a third of the trees identified by the algorithm were false positives. These errors were due to problems with the SLAMalgorithmsused to organize the mobileLiDAR returns into point clouds. A variety of SLAM algorithms were tested, but all had similar problems. We attributed most of the problems with the SLAM algorithms to lack of precision in the data obtained from theinertial navigation systems (INS)(which consistof an Inertial Measurement Unit(IMU) and a GPS). We invested in a somewhat higher quality INS, which improved results, but not enough. Higher quality INSs than ours exist, but were not within the budget for this project.When trees were correctly identified, DBH estimates were typically within 1-5 cm of the correct DBH, with a tendency to overpredict. 3. Develop a hardware/software platform to mount a LiDAR scanner (such as a Velodyne VLP-16) on a unmanned aerial vehicle (such as the DJI Matrice 600 Pro). Our team's original intention was to develop a fully functional aerial liDARsystem with our existing equipment. After a feasibility study, focusing on both time and existing equipment, we concluded this was not an option at present. Integration of an aerial laser scanning system on a UAS platform is a formidable task that requires incredibly sensitive and accurate positional information to be effective. Unlike a ground-based mobile system that can leverage existing SLAM algorithms, an aerial system's narrow field of view relies heavily on inertial navigation systems (INS) to provide accurate positional information to place laser "frames" in coordinate space. This is seen across the existing market offerings, which integrate INS that cost $20,000+ to successfully operate. With our existing equipment, we could not provide both the accuracy and rate of data collection necessary to complete this project. 4. Continue synergistic research activities with PSU Ag and Engineering collaborators. Due to frustrations with development of our mobile laser scanning system, we shifted our focus to utilizing our ground data to develop and test tree identification algoriths using high-quality airborne laser scanning (ALS)data that were obtained for the Stone Valley Forest in 2019 and 2020. Much of the research in individual tree detection (ITD) has involved top-down canopy height model-based methods. These top-down approaches work well in predominately coniferous forests or homogenous settings but perform less favorably in deciduous stands due to the inherent canopy complexity and crown characteristics. Our primary study objective was to develop a scalable individual tree detection method using ALSdata that would perform well in mixed-species hardwood forests such as those found in the northeastern United States. We developed a voxel-based approach that consistently detected 68% of all reference trees greater than 10cm diameter at breast height (DBH) and 87% of sawtimber-sized trees greater than 28cm DBH across 48 1/16th hectare plots. We also developed a new tree-matching method and software that uses linear integer programming to enable the application and validation of spatial accuracy criteria and a more stringent definition of a "detected tree." This approach represents a fundamental change in how ITD methods are evaluated and validated our method's ability to generate accurate tree stem maps with mean positional accuracy of less than one meter.
Publications
- Type:
Theses/Dissertations
Status:
Published
Year Published:
2019
Citation:
Yoder, Kaitlin. 2019. Evaluating New Methods for Predicting Golden-Winged Warbler (Vermivora chryspotera) Site Occupancy Using Unmanned Aerial Vehicle Photography. Penn State University. Master's Thesis.
- Type:
Theses/Dissertations
Status:
Published
Year Published:
2020
Citation:
Hershey, Jeff. 2021. A voxel-based method for individual tree detection using airborne lidar in eastern U.S. hardwood forests. Penn State University. Master's Thesis.
- Type:
Journal Articles
Status:
Under Review
Year Published:
2022
Citation:
Hershey, J. McDill, M.E., Miller, D.A., Holderman, B., and Michael, J.H. In Review. A voxel-based method for individual tree detection using airborne LiDAR in eastern U.S. hardwood forests. Submitted to ISPRS Journal of Photogrammetry and Remote Sensing.
|
Progress 10/01/19 to 09/30/20
Outputs
Target Audience:Forest managers, forestry consultants, people involved in carbon markets, other forest scientists Changes/Problems:We are no longer working on objective 3.We have realized that we do not have the resources (high-end inertial navigation systems) nor the expertise to accomplish this objective. We have decided to focus our efforts on the other three objectives. What opportunities for training and professional development has the project provided?The project provided training for one Master's student, Jeff Hershey, as well as the technician, Brennan Holderman. Of course, we are all learning a great deal as we conduct thisresearch. How have the results been disseminated to communities of interest?We have participated in several meetings with the Department of Conservation and Natural Resources (DCNR) to discusshow the techniques and technologies we are developing can be applied to improve their inventory program. We have also had meetings with representatives of forest products companies,working with the Penn State Office of Technology Management. We have one manuscript in preparation that should be submitted to the journal, Remote Sensing of the Environment, in Spring 2021. What do you plan to do during the next reporting period to accomplish the goals?Goal 1 -After a thorough review of the existing individual tree detection (ITD) algorthims, including 2D raster moving window and watershed on canopy height models, direct point cloud phase shift and ground-up stem modeling, we have determined the incidence of omission and commission amongst complex canopy and heterogenous distribution requires a new approach. With a significant volume of ground-truth data we will have the ability to cross-validate multi-variate models from data derived from our state-of-the-art leaf-on and leaf-off aerial and mobile laser scanning systems. This will allow our team to produce new methods of ITD more tailored to eastern hardwood forest types. Our team will then be able to account for leans, spilt trunks, and crowded stems that are often misrepsented with current approaches. We plan to leverage mobile data to better identify individual stems at the plot level, that will allow for significantly better isolation in forest-level ALS data sets. In addition to the high-density laser data collected from the commercial provider in 2020, Penn State's Office of the Physical Plant contracted another provider in 2019 to collect leaf-off ALS data across the same extent of Stone Valley Forest (SVF). While this data is less dense than the 2020 collection, it will provide the opportunity to evaluate any potential statistical differences in the derived tree metrics and provide insight into the necessary point density for future data collection campaigns. Goal 2 -Our in-house MLS prototype will be deployed to our field sites in Year 2. Without an existing standardized workflow for collection protocols available in the literature, our team will have to evaluate new methods of capture that best sync with ALS data and provide the necessary accounting of tree stems. We intend to evaluate capture methods such as: variable distance transects, standard static plot-level, S-shaped full-coverage, and small area loop closure. The survey we have already completed at our field sites will allow our team to better understand the repeatability of our system and formulate best-practices to mitigate omissions and emphasize efficiency. Goal 3 -No plans forthis objective. Goal 4 -Advances from the continuation of our research have the opportunity to directly influence the future of forest land management in Pennsylvania. In addition, this could be a significant support mechanism to aid in carbon markets, which rely heavily on accurate forest inventory estimations. The work we do has broad applications and the continuation of our efforts provides opportunity for increased synergy to meet the demands of emerging scientific research.
Impacts
What was accomplished under these goals?
Impact: The instruments and data collection and analysis methods we are developing for this project will revolutionize how forests are measured. Instead of using diameter tapes to measure trees on a plot, foresters will use a combination of imagery and Light Detection and Ranging (LiDAR) data from terrestrial and airborne platforms to determine the species composition, volume, value, structure, and biomass of a forested area. The resulting data will be more accurate - virtually eliminating sampling error, while introducing some measurement error - cheaper, and cover far more area than is practical with current methods. This type of information will be useful for forest valuation (for estate valuation, timber sales, land sales, etc.), for forest management planning, and for enabling new markets, such as forest carbon trading. 1. Decades of research in remote sensing for characterizing forest ecosystems and associated metrics have produced a bevy of open-source processing tools. The development of these tools has generally focused on even-aged, or homogenous stands, and their application on more complex structures leave much to be desired. In addition, functions are spread amongst a variety of software libraries and are generally disjointed due the complex nature of processing 3D point cloud and 2D spectral data. To alleviate some of the data processing complexities, our team has focused on distilling the breadth of available tools and libraries into succinct processing pipelines. These pipelines binned into aerial and mobile, and can be considered parallel development projects, as overlap is significant for our purpose. The two pipelines, described in this report under products,are 1) the aerial laser scanning (ALS) data processing pipeline, and 2) the mobile laser scanning (MLS) data acquisition and processing pipeline. 2. The sensors deployed in MLS data acquisition were developed originally for autonomous vehicles. Inherently, this means they were built to move and detect objects in real-time in complex and non-static environments. Several manufacturers have taken these sensors and developed MLS systems. For our team to better evaluate the repeatability and reliability of these sensors, we needed both a gold-standard to compare our results to, as well as test an existing MLS system. We were able to accomplish both of these goals in Summer 2020. 3. Develop a hardware/software platform to mount a LiDAR scanner (such as a Velodyne VLP-16) on a unmanned aerial vehicle (such as the DJI Matrice 600 Pro). Our team's original intention was to develop a fully functional aerial lidar system with our existing equipment. After a feasibility study, focusing on both time and existing equipment, we concluded this was not an option at present. However, our team has successfully developed a mobile laser scanning prototype to fulfill the bottom-up approach to our research endeavors. Integration of an Aerial Laser Scanning system (ALS) on an Unmanned Aircraft System (UAS) platform is a formidable task that requires incredibly sensitive and accurate positional information to be effective. Unlike a ground-based mobile system that can leverage existing Simultaneous Localization and Mapping (SLAM) algorithms, an aerial system's narrow field of view relies heavily on inertial navigation systems (INS) to provide accurate positional information to place laser "frames" in coordinate space. This is seen across the existing market offerings, which integrate INS that cost more than $20,000to successfully operate. With our existing equipment, we could not provide both the accuracy and rate of data collection necessary to complete this project. However, with state-of-the-art traditional ALS data provided by QSI, it is our belief we can successfully meet all other project objectives. 4. Our team has made strides in fostering new relationships across campus for current and future collaborations. The continued collaboration with Brian Nebresenzy and others in the engineering realm will only further our research applications. The technical aspects of our research have garnered much attention from across colleges. Students from aerospace, civil, and mechanical engineering have expressed, or joined, our efforts. Our research is also supporting multiple capstone projects in the Master of Geographic Information Systems (MGIS) progam, as well as in BioRenewable Systems and Forestry Master's programs.
Publications
|