This file was created by scanning the printed publication. Errors identified by the software have been corrected; however, some errors may remain. Industry Perspectives on Implementing and Analyzing an Annual Forest Inventory1 Paul C. Van Deusen 2 Abstract-The USDA Forest Service is moving from a periodic forest inventory to an annual system. The change to an annual system will provide challenges both statistically and logistically. This change has been generally well received by U.S. forest industry for a number of reasons. Periodic inventories served us well in the past, but have failed to keep up with modern information needs. The focus here is on discussing some of the statistical challenges and demonstrating that they can be met with current technology. The conclusion is that annual forest inventory systems make sense in the modern era where current information is essential for meeting regulatory requirements, ensuring sustainability, and providing products from our forests. The USDA Forest Service Forest Inventory and Analysis Program (FIA) is the only source of reliable, national scope information on U.S. forest resources. Historically, FIA has conducted periodic inventories within States on a 10-15 year cycle. In recent years, this cycle has tended to lengthen due to flat budgets and requirements to measure new variables. In the South, pine fiber rotations are shrinking below 20 years, so the periodic inventory is approaching being out of date by a full rotation. While industry inventories its own land, the majority of the fiber supply for most mills does not come from company owned lands. Concern about FIA was formally expressed in 2 Blue Ribbon Panel (BRP) reports written by forestry experts from industry, academia, government and environmental organizations. Among other things, the first report (BRP 1) in 1992 called for a decrease in the FIA survey cycle to 5 years. Some short-lived progress was made in this direction, but FIA cycles were of unprecedented length 5 years after BRP 1. A second report (BRP II) came out in 1998 including statements such as "the lack of major improvement in FIA is leading to the loss of important ecologic and economic benefits to society by hindering our ability to monitor forest health and sustainability." BRP II called for initiating an annual inventory system where a proportion of plots would be measured every yearin every State. Following BRP II, the Research Title of the 1998 Farm Bill mandated that FIA move to an annual system where 20% of the FIA plots are measured annually in each State. The move to an annual system is a momentous event for FIA that rivals any other change since FIA's inception. Ipaper presented at the North American Science Symposium: Toward a Unified Framework for Inventorying and Monitoring Forest Ecosystem Resources, Guadalajara, Mexico, November 1-6,1998. 2Paul C. Van Deusen is Mathematical Statistician, National Council of the Paper Industry for Air and Stream Improvement (NCASI, Inc.), Tufts University, Department of Civil Engineering, Medford, MA, U.S.A. Phone: (617) 627-2228; Fax: (617) 627-3831. e-mail: pvandeus@tufts.edu; http://NCASIl.nerc.tufts.edu:443 230 Such a change entails risk, but, from forest industries view, the change is needed. There are several reasons why an annual system makes good sense. 1) Politically, annual systems will allow FIA to maintain their contacts within a State, which is hard to do with the 10 year hiatus that currently exists. 2) Budgetarily, regular annual budgets will enable greater participation by the States, i.e. it's difficult to generate a budget at irregular intervals. 3) Informationally, the annual system will provide current data that meshes seamlessly across States, which is not a characteristic of the current system. The value of current information will be the focus of the remainder of this paper, since that is what motivates industry to support an annual system. Some comparisons will be made between capabilities afforded by annual versus periodic systems to: 1) provide estimates of means, 2) provide estimates of change, 3) detect trend changes, and4) incorporate remote sensing technology. Mean Estimates -------------------------Periodic surveys measure all N plots in a short amount of time, ideally over 1 year. An annual inventory taking a 20% sample measures N/5 plots each year. Therefore, a naive analysis would suggest that confidence intervals for the annual inventory estimate are ~5 times larger than for the periodic estimate. This result assumes that only the current year's 20% sample is used to compute the desired mean. This is an unfair comparison since periodic surveys of a State often take 2 or 3 years to complete and estimates are then based on averaging all plots together. Following this same approach, annual inventories can use moving average estimators over the current and previous 4 years to derive estimates that will have confidence intervals nearly as narrow as from the periodic survey. An important caveat is that the periodic survey only provides an estimate roughly every 10 years leaving the user to interpolate for intermediate years with unknown confidence. Forests grow slowly, and previous year data contain nearly as much information about current forest characteristics as current year data, unless a major disturbance has occurred. Likewise, data from several years earlier still contain valuable information. A well constructed moving average estimator can down-weight the earlier years to take advantage of this information. Moving average estimators can be constructed in a number of ways. For example, Van Deusen (1998) demonstrates how to use a mixed estimator, which is closely related to the Kalman filter. There are other approaches to analyzing annual data that can improve upon the naive approach as well. For example, unmeasured plots could have their current values imputed or modeled, and these imputed values could be incorporated USDA Forest Service Proceedings RMRS-P-12. 1999 into the estimate. With very accurate imputation, the results of this approach could allow one to obtain results that are nearly as good as if all plots were measured annually. Statistical methods for dealing with imputed or modeled values include multiple imputation and double sampling (Fairweather and Turner 1983, Rubin 1987, Hansen 1990, Van Deusen 1997, Reams and Van Deusen 1998). Change Estimates _ _ _ _ _ __ Change estimates from periodic surveys indicate the difference between current values and those of 10 years earlier. The disadvantage here is that there is no way to know if the change occurred smoothly throughout the period or in some irregular manner. Annual inventories allow for annual change estimates, which supports a more sensitive ability to monitor trend. This is generally a positive attribute, but could lead one to react prematurely to temporary trends or artifacts of sampling error. In spite of the danger of premature reaction, changes in trend should be detectable sooner with an annual inventory than with a periodic system. Trend monitoring will be facilitated by an annual system because the development of a trend can be followed from one year to the next. However, there is currently a need for research in this area. Statistical methods need to be proposed for testing the null hypothesis of no-trend versus the alternative of increasing or decreasing trend. Studies are needed to determine the impact of changing cycle length, i.e. measuring a different proportion of plots each year, on the power of these tests. This will be important to weed out the irrelevant blips that will inevitably occur if simple annual means are plotted. In fact, it would be very surprising if raw sample means from an annual inventory produced a smooth trend. Remote Sensing Traditionally, the FIA program has produced area estimates of forest and non-forest using double sampling. The process consists of interpreting a large number of sample points on aerial photographs and subsampling a proportion of the points on the ground. Complete forest type maps would be difficult to produce from photography due to the tremendous amount of manual interpretation required. The annual system's requirements for regular generation of land-use and land-cover maps would be prohibitively expensive with aerial photography. Satellite based imagery provides the basic data needed to classify cover-types of large areas in an automated, cost effective and timely manner. The three most readily available (at this time) satellite sensors are the thematic mapper (TM), SPOT, and AVHRR. Presently, TM data is preferred because TM has greater spectral resolution than SPOT, and better spectral and spatial resolution than AVHRR. To estimate map class area totals and variances FIA may be able to use two-phase or double sampling where the less accurate data is the map whose accuracy is in question, and the more accurate but costly data is the FIA ground sample. A sampling scheme designed to evaluate and correct for map area misclassification is as follows: A sample of n points/pixels is located on the map and the true and map USDA Forest Service Proceedings RMRS-P-12. 1999 categories are determined for each point. The n points are allocated as a simple random sample. This results in a two way contingency table where nij is the number of points in the sample whose true category is "i" and whose map categoryis "j". Formulas for estimating the true probabilities of interest are given in Card (1982) along with variance estimates. Methods for estimating change in category proportions between two times are given in Van Deusen (1994) along with variance formulas. Coincidentally, the estimators for the true map proportions are the same for simple random sampling (srs) or stratified sampling of map pixels. However, variance estimators are different under the two sampling strategies. Judicious use of remote sensing can provide benefits to the annual system beyond the ability to produce complete cover maps. Efforts to model unmeasured plots will be aided if the disturbance status of the plot is known. The USDA Forest Service's North Central region is already conducting research on disturbance detection. Likewise, better modeling capability would allow for reducing the number of ground samples while maintaining estimate precision. Discussion ------------------------------Changing from a periodic inventory to an annual system involves some risk. However, annual inventories have been under study by the North Central FIA since 1992 and in the South since 1995. The North Central annual inventory system (AFIS) was based on measuring disturbed plots with a higher probability than undisturbed plots. The southern system (SAFIS) was based on measuring an equal proportion annually with systematic coverage. The AFIS design may have some advantage from an efficiency perspective, but it is more difficult to statistically analyze. The SAFIS design was selected because of its simplicity and robustness. This reduces the risk. The 1998 Farm Bill called for implementing a SAFIS-like design for this reason. An inventory based on regular annual samples differs from the periodic survey primarily in the timing of plot visits. The same plots are used and the same measurements are made under both systems. Therefore, the biggest uncertainty lies in logistical issues. If the logistical problems can be solved then the annual inventory will succeed. Experience so far in States like Minnesota, Virginia and Georgia indicates that the new system will be a success. Even though logistics will create the initial hurdles, research on statistical analysis and remote sensing will be required to get the most out the annual data. Initial procedures to analyze the data and to implement remote sensing should be ready when enough data to begin assessing trends are available. The annual systematic design lends itself to several analysis options and to remote sensing methodologies as well. The robust design will allow for adaptive improvement of the analysis techniques as experience with the system is gained. I see little downside to an annual inventory system relative to the periodic system. It has already been demonstrated in several States that the data can be efficiently collected. However, the selection of an analysis method has not been finalized. Confidence interval width will depend on the analysis method used, so it may be some time before a 231 comparison with the periodic system is complete. This comparison is also complicated by the fact that there is no estimate or confidence interval for intermediate years for a periodic inventory. The data collected under an annual inventory system are identical to what was collected under a periodic system, unless changes are made for unrelated reasons. Therefore, database management will be quite similar except for the fact that annual data will be continuously arriving. Users will want data access more frequently because hard copy reports will not be released annually. Industrial users, in particular, will want to access the data and perform custom analyses on a regular basis. It would be ideal if software were available on the internet to both access and analyze these data. In fact, FIA has already made progress in this direction and should be congratulated, but more is needed. Industrial users view FIA as providing the only data that are broad in scope and focused on the forest resource. FIA data are critical to assuring long term forest sustainability in which forest industry has a vested interest. The data that it will produce are so important that failure of the annual system is not an option. 232 Literature Cited Card, D.H. 1982. Using known map category marginal frequencies to improve estimates of thematic map accuracy. Photogrammetric Engineering and Remote Sensing. 48(3):431-439. Fairweather, S.E. and B.J. Turner. 1983. The use of simulated remeasurements in double sampijng for successive forest inventory. In Proceedings, Renewable Resqurce Inventories for Monitoring Changes and Trends. August 15-19, 1983, Corvallis, Oregon. John F. Bell and Toby Atterbury, editors. Pg. 609-612. Hansen, M.H. 1990. A comprehensive sampling system for forest inventory based on an individual tree growth model. PhD dissertation. Univ. of Minnesota. St. Paul, Minn. Reams, G.A. and P.C. Van Deusen. The Southern annual forest inventory system. IN PRESS. Journal of Ag. BioI & Env. Stat. Special issue: Environmental Monitoring Survey Over Time. Rubin, D.B.1987. Multiple Imputation for Nonresponse in Surveys. Wiley. Van Deusen, P.C. 1994. Correcting bias in change estimates from thematic maps. Remote Sens. Environ. 50:67-73. Van Deusen, P.C. 1997. Annual forest inventory statistical concepts with emphasis on multiple imputation. Canadian Journal Forest Research 27:379-384. Van Deusen, P.C. 1998. Alternative sampling designs and estimators for annual surveys. IN: Proceedings of International Conference on the Inventory and Monitoring of Forested Ecosystems. Boise, ID. August 1998. Eds. M. Hansen and S. Fairweather. USDA Forest Service General Tech. Report. USDA Forest Service Proceedings RMRS-P-12. 1999