Planificare strategica forestiera la nivel de arbore folosind technici de teledetectie (RemoteForest)

 

Cod proiect: PN-II-PT-PCCA-2011-3.2-1710


Prezentare proiect

Finanțat de UEFSCDI

Unitatea Executiva pentru Finanțarea
Invatamantului Superior, a Cercetării,Dezvoltării si Inovării

www.uefiscdi.gov.ro

Institutii Partenere

ICAS Bucuresti – Institutie Coordonatoare
INCD Stiinte Biologice – Partenerul #1
Universitatea Stefan cel Mare – Partenerul #2
Ocolul Silvic “Codrii Beiusului” – Partenerul #3: Partener din industrie

Personal

Bogdan Strimbu
Gheorghe Marin
Andrei Paun
Marian Dragoi
Ferko Jano

Summary

The strategic forest planning process aims at optimization of a series of objectives that are subject to a set of constraints. The incorporation of market dynamics in the optimization increases the complexity of the strategic planning formulation by requiring the inclusion of detailed tree level products (such as sawlogs, pulpwood, or chip ‘n saw) in the planning process. Furthermore, constant adjustment of planning objectives to short-term market changes requires fast and accurate identification of the products that can be supplied by the forest estate subject to strategic planning. The advent of remote sensing techniques, especially LIDAR, reduces the time to acquire accurate information required for products allocation. The present research aims at identification of products allocation based on LIDAR information that would supply the optimal solution to the strategic planning objectives. LIDAR data will be used to determine a series of tree attributes (e.g., total height, crown width, crown length and crown asymmetry) that will be used to delineate possible products to be obtained from a tree. Models describing each tree in terms of products allocation will be developed by adjusting taper equations to the attributes estimated using LIDAR. The optimal products allocation would be determined at the forest estate level using several planning algorithms, including linear programming, simulated annealing, and first fit decreasing algorithm constrained to fulfil the perfect bin-packing theorem. The research will provide the information needed by Romanian forest operators to adjust their strategic and tactical planning to market conditions.

Concept and rational of the project

Concept of the project

The main concept of the project is the integration of forest planning and evaluation services with market demand for forest related products (forests cover 27% of the country, with less than half having a dedicated production objective). The project will evaluate and predict the forest inventory using remote sensing (LIDAR and aerial orthophotos), a novelty for the forest operators in Romania. A major novelty would be related to the correlation of the inventory to the market changes, therefore, leading to improved forest management decisions.
The project is a partnership between leading Romanian scientists acting in quantified silviculture areas and one of the largest private forestland holders and processors in Romania (Ocolul silvic Codrii Beiusului). To process the large amount of data generated by remote sensing techniques the LONI network will be used (to which the coordinating institution (i.e. , ICAS) has a formal memorandum of understanding), for the benefit of a large part of Romania’s agents actively involved and operating in the management of environment or renewable natural resources.
The tools developed and the results obtained will be disseminated widely to forest operators, timberland owners as well as to private or public entities acting in environmental area.

Rationale of the research

A successful entity that manages forested land has accurate knowledge of its forest inventory, not only in respect to the amount available, expressed in tons, cubic feet or board feet, but also in respect to the type of products that can be obtained from the available resources. The inventory provides a momentary description of the magnitude and the possible products that can be obtained from the forest, but it has little connection with the present or future economic and social needs. To couple market dynamics with changes in forest inventory complex, forest plans are developed and solved aiming toward the fulfilment of the Sustainable Forest Management (SFM) criteria. However, the mathematical models that are used in forest planning do not necessarily supply the optimal solution as they face two challenges: 1) large amount of data to be processed, and 2) difficult to implement social, economic and environmental constraints in real time. The present research addresses both difficulties by developing and implementing a method that integrates fast and accurate resource inventory tools, based on LIDAR and aerial orthophotos, with market dynamics, using forest planning.

Goals and Objectives

The goal of the present research is the identification of products allocation supplying the optimal solution to the strategic planning problem using LIDAR. The research would use forest inventory information determined by integrating LIDAR with aerial ortho-photo images. Tree attributes derived from LIDAR data (e.g., total height, crown width, crown length and crown asymmetry) will be used to identify possible products that can be obtained from each tree. Tree level products will be delineated by adjusting taper equations to the attributes estimated from LIDAR. The optimal products allocation for the entire forest estate would be determined using several algorithms, commonly implemented in forest planning.

To ensure the achievement of project’s goal a series of annual objectives should be reached. The objectives are serial and summative. The objectives for the intervening years are:

First year:

  1. acquire the remote sensing data (LIDAR and aerial ortho-photos),
  2. identify tree products,
  3. start the sampling design, and
  4. initiate the planning procedures on the supercomputer LONI (Louisiana Optical Network Initiative).

Second year:

  1. completion of remote sensing data acquisition, sampling design, and field measurements,
  2. development and completion of the taper equations used to identify products allocation,
  3. completion of all the inputs needed for forest planning, namely product allocation at tree and stand level as well, as the additional constraints fulfilling social, economic and environmental requirements.
  4. develop PC software to be used by forest operators that do not have access to high computational devices, and
  5. write a report presenting the sampling methodology for LIDAR estimates

Third year:

  1. determine the forest planning solution using LONI and the PC software,
  2. identify and assess the impact of products allocation on the forest planning process, and
  3. write at least two peer review article, one comparing the results obtained using LONI and PCs, and one presenting the response of strategic planning to product allocation.

A direct result of the project would be a significant decrease in the inventory costs and the shortening of the interval between two inventories, two major components of the adjustment to market dynamics of forest industry. It is expected that successful completion of the project would polarize the regional forest planning research and the remote sensing research performed to northern Louisiana. It is also expected that a spin-off company will be created to address the remote sensing processing needs of regional forest operators or other entities acting at landscape level. Furthermore, a formal sampling and landscape planning laboratory could be created at Louisiana Tech University. The laboratory coordinated by two professors (Dr. Strimbu and Dr. Paun) will include scientists from the Forestry, Computer Science and Statistics departments, and would have at least three graduate students. Furthermore, it is expected that an increase with 15% in the personnel of the School of Forestry would occur in the next 5 years due to the successful completion of the project.


Methodology

LIDAR data acquisition

The main source of data for this research would be obtained using airborne LIDAR, which provides direct measurement of elevation of different parts of the tree. The parts of interest are the ground (base of the tree), top and base of the crown. The knowledge of these three values allows the identification of total tree height, height to the base of the crown and the length of the crown. Depending on the sensor used to acquire LIDAR data, the information used for determining the previous three attributes can be continuous or discrete (Gatziolis and Andersen, 2008; US Census Bureau, 2009). The study will use a discrete return system to acquire LIDAR data (Evans et al., 2009), as it can supply tree height with an accuracy similar to ground measurements. The multiple returns feature of LIDAR pulses allows also an accurate identification of the forest floor, which is translated in an increase in the accuracy of the vertical attributes (Evans et al., 2009). To ensure the correct identification of the tree, the pulse density would be at least 6 pulses / m2, as recommended by Evans et al. (2009) or Gatziolis and Andersen (2008), and similarly to Andersen et al (2006). The relatively high density of pulses would allow the determination of stand height with an expected accuracy of less than two feet. The overlap of flight line would be at least 50% to allow scanning of the trees from multiple look angles, which supplies a more complete 3-dimensional representation of the trees (Evans et al., 2009). The flights would be executed only during the leaf-off period, which in Louisiana is commonly without snow. The selection of the leaf-off season for flights is recommended by the increased probability of the LIDAR sensor to receive pulses that reached the ground and tree stem, which is larger than during the leaf-on period. Other attributes, such as diameter at breast height, diameter at the base of the crown, and presence/absence of bifurcated trunk, which can be used to identify the products that can be obtained from each stem, will be determined using regression equations based on the tree attributes derived from LIDAR data. Species, which plays a critical role in identification of the products that can be obtained from a stem, is difficult to be identified using only LIDAR data. However, a significant number of procedures were developed to identify the species, procedures that combine LIDAR data with other remote sensing acquired data, such as QuickBird images (Ke et al., 2010), hyperspectral data supplied by AVIRIS (Airborne Visible Infrared Imaging Spectrometer) and LVIS (Laser Vegetation Imaging Sensor) (Swatantran et al., 2011), or airborne digital ortho-photo images (St-Onge and Achaichia, 2001). The project will utilize ortho-rectified aerial photographs obtained using multispectral airborn dedicated flights with accuracy of 0.5 m.

Ground data measurements for validation and calibration

The achievement of the main objectives of the project is significantly influenced by the details associated with the forest inventory and forest planning process. To ensure that LIDAR-aerial orthophotos derived estimates are accurate, tree level information (e.g., total height, height to base of the crown) will be verified using 1/5 acre plots. The trees that are located inside the plots will be identified using the maximum filtering algorithm developed by Popescu et al (2003) within the LIDAR data, while in the field the position of the trees will be obtained using GPS units and Haglof DME 201 cruisers. The number of plots needed to describe individual tree geometry would be determined by using the LIDAR estimate alone (namely the coefficient of variation, or mean and standard deviation) and then aggregating the trees within the plots. The estimates obtained by processing only LIDAR data, without any ground verification, would be biased, but for sampling purposes it is not the magnitude of the attribute that is of interest but the variability existent within the data. It is assumed that the coefficient of variation, the statistics representing data variability used in sampling computations, is accurately quantified using LIDAR data, even that the mean and standard deviation are actually biased. The main sampling technique used to acquire ground information would be 3P sampling within a stratified random sampling framework, as recommended by Cochran (1977). The thresholds for the elements used to determine the sample size (i.e., sampling error and probability to reject null hypothesis) would be consistent with the timber cruising handbook of the US Forest Service – tree measurement scale (Robertson, 2000). The stratification will be executed in two directions structured in a factorial manner: 1) a stratification will separate the area based on species or species groups (such as pines, hardwoods or mixed species), and 2) a stratification will separate the area based on site productivity (i.e., low, medium, and high). The selection of strata were recommended by the accurate identification of the products, which are species dependent, and by the change in taper with site productivity (Clutter et al., 1983). A minimum of 90 plots will be measured (i.e., 3 sites x 3 species x 10 plots). The usage of fixed area plots would allow also the determination of stand level estimates, to be used in the forest planning. For all trees with DBH>4 in inside each plot the following attributes will be recorded: total tree height, height to the base of the crown, DBH, the products that can be obtained, and other issues identified on the stem (such as bifurcation, defects induced by insects or fire). Tree location would be mapped by recording the coordinates for each plot centre using GPS units, and azimuth and distance with respect to plot centre of every tree. Considering the positional accuracy of the differential GPS unit for determining the location of plot centres, the horizontal location error of a tree is expected to be less than 7 ft (approximately 2 m). This error only refers to the position of the base of the tree, without considering the deviation of the tree top relative to the base (Popescu, 2007).

Tree and stand level allocation of products

The products that can be obtained from a stem will be determined using taper equations, as presented by Nunes et al. (2010). Based on the results obtained by Newberry and Burkhart (1986), Jiang and Liu (2011), and Shaw et al (2003), the taper functions will use as input the estimates supplied by the combination of LIDAR-aerial ortho-photos, at least species, height and crown ratio. To determine the total height, crown diameter and height to live crown for each tree the automated algorithm presented in several papers (Popescu et al., 2002, 2003) and implemented in TreeVaW (Kini and Popescu, 2004) will be used. It is expected that products identification will be executed for approximately 5 million trees. The massive computation will be performed on LONI.

The expansion from tree level products to stand level products will be done by adding the estimates of all the trees within the stand. This approach, advocated by Popescu et al (2004), is preferred, as it does not require the identification of the stem (a process that is computational intensive, and still does not have an accepted algorithm). Furthermore, the simple addition of tree level estimates is parsimonious, fast, and provides accurate results (Popescu et al., 2004).

The field measurements will be compared with the estimates derived from the combination LIDAR- aerial photographs, and will serve a dual purpose: 1) to assess the accuracy of LIDAR estimates, and 2) to develop equations for the yield of each product potentially to be obtained. The equations will use as input variables the values supplied by LIDAR and as predicted variables the yield of a product that can be obtained from a particular tree, and consequently, from the stand in which the tree belongs.

Strategic forest planning aiming optimal product allocation and SFM

The individual tree constitutes the elementary unit of the strategic plan designed to supply the optimal amount of raw material to industrial processes yielding products synchronized with market demand. For the selected area, it is expected that approximately 5 million trees would be included in the planning process. The large number of trees combined with the constraints framing the SFM (such as green-up adjacency, wildlife corridors, or biodiversity conservation) will require complex mathematical algorithms to solve the optimization problem. The economic constraints will consider not only continuity of the existing processing facilities (an SFM requirement) but also the stock market volatility, namely the variation of the main stock market indexes (such as Dow Jones or S&P 500), and the value of the products that can be obtained from a tree, as evaluated by the New York Mercantile Exchange. A set of four heuristic algorithms will be implemented to identify the optimal product allocation for strategic planning: simulated annealing (Lockwood and Moore, 1993), genetic algorithms (Bettinger et al., 2002), tabu search (Cvijovic and Klinowski, 1995), and first-fit decreasing algorithm (Strimbu et al., 2010). The implementation of the planning algorithms will be performed on LONI, which can process the massive computational planning problem proposed.

The large number of trees makes the planning strategy using tree level information on a PC impractical. To ensure the applicability of the project’s findings to entities that do not have access to high computational facilities but only high performance PCs, a reductionist alternative will be developed based on the results supplied by LONI. The reduction consists in grouping the trees from the same stand that have similar products obtained from their stem. The grouping will create a new elementary planning unit (i.e., an entity grouping several trees), which will decrease the size of the problem by an expected two orders of magnitude (from 5 million to approximately 50,000). The PC elementary planning unit is either a portion of a stand (not necessary contiguous), or the entire stand on which the trees grow. At least five sets of products for each elementary planning unit would be considered, which would constitute the basis for identification of the optimal products allocation. The implementation of the strategic planning on PC will be executed by enhancing one of the three software: Atlas (produced at the University of British Columbia), Janus (produced by Louisiana Tech University), or SiMO (. Both software packages are spatially and temporally explicit but differ in the implementation algorithm, one based on sorting according to different criteria, such as stand age or distance between stand and access road (Atlas), and one based on simulated annealing (Janus). The usage of multiple analytical platforms was advocated by Duinker and Kennedy (1994), and implemented by Strimbu and Innes (2011), who proved that a unique framework does not provide valuable forecasts, especially in strategic planning, at most possible but not necessarily likely future (i.e., probability of a future event to happen is asymptotically null).

A large number of researchers proved that heuristic techniques are sensitive to the details describing the algorithm (such as the initial temperature and annealing rate in simulated annealing), and the planning problem itself (such as harvest age or green-up adjacency delay). The present project will assess, in addition to identification of optimal product allocation, the sensitivity of optimal product allocation to the details of the planning algorithm using a mixed model repeated measures analysis organized as a factorial design. The factors considered in the investigation are the variables characterizing globally the planning algorithm, such as time to produce an optimal solution, number of harvests of an elementary planning unit during the planning period, or minimum distance to roads. The repeated measures analysis considers the recurring evaluation of the product allocation obtained from the investigated area, which is imposed by the strategic planning.

LONI-LIDAR estimates and product allocation

LONI is a state-of-the-art, fibre optics network that runs throughout Louisiana, and connects Louisiana and Mississippi research universities to one another as well as National LambdaRail and Internet2. LONI provides Louisiana researchers with one of the most advanced optical networks in the USA and the most powerful distributed supercomputer resources available to any academic community with over 85 teraflops of computational capacity. Queen Bee, the core supercomputer of LONI and the 23rd fastest supercomputer in the world according to the June, 2007 Top500 listing, is a 50.7 TFlops Peak Performance, 668 compute node cluster running Red Hat Enterprise Linux version 4 operating system. Each node contains dual Quad Core Xeon 64-bit processors operating at a core frequency of 2.33 GHz. Queen Bee was housed at the state’s Information Systems Building (ISB), Baton Rouge.


Results

ISI journals

  • Strimbu, B.M. and Paun, M. (2013) Sensitivity of forest plan value to parameters of simulated annealing. Canadian Journal of Forest Research 48(1): 28-38
  • Strimbu BM 2014 Comparing the efficiency of intensity-based forest inventories with sampling-error-based forest inventories. Forestry 87(2): 249-255 (IF=1.865)
  • Andrei Paun, Petr Sosík: Three Universal Homogeneous Spiking Neural P Systems Using Max Spike. Fundam. Inform. 134(1-2): 167-182 (2014) (IF=0.479)

Conferences

  • Strimbu, V.F. and Strimbu, B.M. Product identification of standing trees using hemispherical photography. ForTechEnvi – Forest and Forest Products Technology and the Environment, Brno, Czech Republic (27-31 May 2013 )
  • Palmer, W., Nitu., D.M., Cooke, D.B., Strimbu, B.M. Accuracy Estimation of Attributes Describing Forested Landscapes Using Tree Level Information. 2012 Annual Symposium of the US – International Association for Landscape Ecology. Austin TX (Apr 14-18, 2013)
  • Strimbu, B., Cooke, D. B., Strimbu, V. F. 2013 Louisiana Academy of Science Annual Meeting, Louisiana Academy of Science, Grambling, LA, “Deterministic algorithm for tree detection from LIDAR data”, (March 9, 2013).
  • Strimbu, B., Cooke, D. B., Strozier, S. 17 Biennial Southern Silvicultural Research Conference, USFA and Louisiana Tech University, Shreveport LA, “Sampling for compliance with USFS guidelines using information derived from LIDAR and multispectral aerial photography”, (March 6, 2013).
  • Andrei Paun, Manuela Sidoroff, Small Universal Homogenous Spiking Neural P Systems Using Max Spike, Annals of University of Bucharest, section Computer Science, DACS 2014 at 11th International Colloquium on Theoretical Aspects of Computing, pp. 79-96 (conferinta ERA de categorie B)

Referate științifice