Improving Forest Inventory: Integrating Single Tree Sampling With Remote Sensing Technology C.J. Goulding 1, M. Fritzsche 1, D.S. Culvenor 2 1 Scion, New Zealand Forest Research Institute Limited, Private Bag 3020, 49 Sala Street, Rotorua 3046, New Zealand. tel +64 7 343 5899 chris.goulding@scionresearch.com, marika.fritzsche@scionresearch.com 2 CSIRO, Private Bag 10, Clayton South, Victoria 3169, Australia Darius.Culvenor@.csiro.au Abstract Individual tree sampling is an alternative method to plot-based inventory for forest assessment particularly in small, irregularly shaped areas. The biggest obstacle to adoption by industry is obtaining an accurate total tree count with an unbiased method of selecting and measuring sample trees. If this could be achieved then per hectare sampling variation across the stand could be avoided and sampling need only be concerned with between tree variability. In order to develop a cost-effective procedure, Scion and CSIRO are testing semi-automated estimates of tree counts from remotely sensed imagery combined with methods of ground-based sample tree selection and measurement. TIMBRS is a user-friendly, tree identification and crown delineation program that analyses high spatial resolution digital imagery in a top down, spatial clustering approach using local spectral maxima and minima to delineate crown centres and boundaries. Measuring individual trees on the ground requires that the sample trees be selected randomly in an unbiased manner without incurring excessive walking time and cost. Boundary trees must be appropriately represented. A very narrow transect laid out in a zig-zag or a structured walk achieves this with minimal bias while reducing non-productive time, almost eliminating difficult walking through thick vegetation where the stand abuts open land. Validation of the project concept has been completed, with good estimates of stocking. Double sampling, where every tree encountered in the walk is measured for dbh and a proportion cruised to estimate merchantable log-grade mix, improved sampling efficiency and demonstrated that the concept was viable. Introduction The intensively managed plantations of New Zealand and Australia are measured using in-place inventory, several times throughout a stand s rotation. The Radiata pine stands yield multiple log products with a wide range of grades and processing destinations. Information from merchantable volume inventory obtained at mid-rotation and pre-harvest is used to optimise the value to be recovered from a stand. The current operational system assesses stem-wood qualities by cruising individual trees measured in sample plots accessed from the ground. The per hectare values are then multiplied by estimates of net stocked area to derive totals, (Deadman and Goulding, 1979; Gordon, Lawrence and Pont, 1995; Gordon, Wakelin and Threadgill, 2006). A typical company would inventory area s averaging 40 to 50 hectares or less, with a plot density of one to every one or two hectares. If an accurate total tree count could be obtained from remotely sensed imagery, individual tree sampling could provide an alternative method to plot-based inventory, avoiding per hectare
sampling variation across the stand. Sampling need only be concerned with between tree variability, reducing field costs. In small stands with irregular boundaries, such as farm woodlots, the need for an accurate estimate of net stocked area is avoided. Image Processing for Tree Identification TIMBRS is a user-friendly, semi-automated implementation of the TIDA tree identification and delineation algorithm, originally developed for Eucalyptus forests (Culvenor, 2002, Culvenor et al., 1998). The spatial resolution of imagery required for accurate tree counts depends on the type of forest and its age class. Satellite imagery from the QuickBird high spatial resolution sensor was available for the study area. The imagery has approximately 2.4 m spatial resolution in its multi-spectral bands (blue, green, red and near-infrared) and 0.6 m spatial resolution in a single panchromatic band. Prior to importing the images into the software, the multi-spectral bands were spatially sharpened to an effective spatial resolution of 0.6 m using the panchromatic band. Within a stand defined by a polygon delineated on the image, TIMBRS used local spectral maxima as indicators of the likely location of a tree crown. Three small stands were first assessed using TIMBRS and then the actual total stocking in each of the stands was obtained by physically counting all the stems in the field. Fully automated tree location and counting without any user interaction may be highly desirable, but this cannot yet be achieved with any degree of confidence. Seen from above, Radiata pine crowns are unruly, sometimes with multiple leaders, irregularly shaped crowns and broken tops, as well as the problems of determining the crowns of suppressed trees. In this study, all semi-automated image processing was carried out by a forestry scientist with no previous experience who, once familiar with the software, could produce an acceptable result within an hour, see Figure 1 and 2 and Table 1. Figure 1. QuickBird image of Stand. Figure 2. Acceptable tree count.
Site Stand Age Area Field Tree count TIMBRS Tree count Comparison 1 106/2 27 7.5ha 1865 1846 99% 2 849/1 30 4.6ha 1691 1691 100% 3 893/1 32 4.4ha 974 1025 105% Individual Tree Sampling Table 1. Accuracy of tree counting by image processing Selecting individual trees in an efficient manner such that they constitute a valid random sample is not simple in practice. The aim was to develop an auditable field method that was practical, reduced walking time and could be used by commercial field crews. After some experimentation, it was decided to traverse the stand in a set of narrow transect paths, measuring each tree encountered for dbh and additionally cruising alternate sample trees for stem characteristics in order to estimate log product yield. Trees encountered were approximately 10 to 12 m apart, see Figure 3. The three trial stands had been measured prior to the project using existing practices by regular inventory contractors to provide a control. Circular plots of 0.02 to 0.04 ha had been laid out on systematic grids. The stands varied in area from 4 to 7.5 ha, with low to high hindrances due to understorey and steep terrain. In two of the stands, a boundary was ill-defined in the GIS and the net stocked area was poorly estimated. Individual tree sampling was trialled in each stand, aiming to measure approximately 100 trees for dbh, half of which were also cruised for merchantable volumes as a double sampling option. Figure 3. Individual trees and samplelines.
Sampling Efficiency Figure 4 shows the change in the confidence intervals expressed as a percentage of the mean merchantable volume (total recoverable volume, TRV) per hectare for varying numbers of sample trees measured in plots or separately as individual stems. For equivalent confidence intervals, single tree sampling required far fewer stems to be cruised than sampling using bounded plots. The same trends were apparent for the estimates of volumes for each of the log grades. Figure 4. The Change in Precision of Total Recoverable Volume (TRV) with Sample Size: Plot- and Individual Stem- sample based. Conclusions Single tree sampling could be more efficient in estimating merchantable volume than conventional bounded plots, given an accurate tree count from remote sensing. The image processing software does not require specialist operators, but does require some knowledge of the forests, as would be expected of a resource forester. Field costs are directly related to the amount of walking through the stand, the time to layout a circular plot and the number of trees to be cruised for stem qualities. All three cost components were significantly reduced using individual tree sampling. The map of individual tree locations is useful in its own right. A key requirement of the field method is that each tree be selected in an unbiased manner with approximate equal probability and with minimal correlation between adjacent sample trees. An alternative sampling design that reduced the delays caused by dense vegetation hindrance on the stand border is to lay out the transect paths in a zig-zag pattern, or to use a pre-planned
structured walk where the field-crew began and ended the sample lines at the same, convenient point (see MacLaren and Goulding, 1993). Further work is being carried out to confirm that total costs including imagery and processing are less than those incurred by current inventory practice. Literature Cited Culvenor, D.S., 2002. TIDA: An Algorithm for the Delineation of Tree Crowns in High Spatial Resolution Remotely Sensed Imagery, Comput. Geosci. 28 (2002), pp. 33 44. Culvenor, D.S., Coops, N.C., Preston, R. and Tolhurst, K. 1998. A spatial clustering approach to automated crown delineation. In, Hill, D.A. and Leckie, D.G. eds. Automated interpretation of high spatial resolution digital imagery for forestry. Victoria, British Columbia, pp 67-80. Deadman, M.W. and Goulding, C.J., 1979: A Method for the Assessment of Recoverable Volume by Log-types. New Zealand Journal of Forestry Science 9(1): 225-39. Gordon, A.D., Lawrence, M.E. and Pont, D. 1995. Assessing the Potential Log Yield of Stands Prior to Harvesting. In Proceedings of the Institute of Foresters of Australia 16th Biennial conference "Applications of New Technologies in Forestry." Ballarat, Victoria. 18-21 April 1995. Gordon, A.D., Wakelin, S.J., Threadgill, J.A., 2006: Using Measured and Modelled Wood Quality Information To Optimise Harvest Scheduling And Log Allocation Decisions. New Zealand Journal of Forestry Science 36(2/3): 198 215 (2006). MacLaren, P. and Goulding C.J. 1993: The structured walk - a practical inventory system. New Zealand Journal of Forestry 37(4):20-23.