Towards site index mapping in deciduous stands using multi-echo LIDAR data
J. Bock1, E. Dambrine2, G. Dez1,3, J.-L. Dupouey3, M. Georges-Leroy4, A. Jolly1,
F.S.R.V. Martins1,3 and J.-P. Renaud1
1. Office National des Forêts, Nancy, France;
2. Forest Ecosystems Biogeochemistry, INRA-Nancy, France;
3. Forest Ecology & Ecophysiology Unit, INRA-Nancy, France;
4. Servicerégional de l'Archéologie de Lorraine, Metz, France;
Introduction
Site fertility index is an important parameter for forest management. Its calculation requires the knowledge of stand age and dominant height (Ho) which are both difficult to acquire. Whereas age could be deduced from plantation dates found in archives, measuring Ho is often tedious. Therefore, site index is rarely mapped over large areas with a fine scale resolution. Ho is also in itself of practical importance for forest management since it is a key parameter involved in planning thethinning operations.
Airborne LIDAR systems are now used operationally for forest inventory at various geographical scales (Naesset 2002). However, to estimate parameters of interest, no generic equations are available owing to variabilityin data acquisition systems, standcomposition and structures. Therefore,the use of LIDAR metrics requires the generation of empirical calibrations equations from ground-truth plots.
Several sources of variability may affect accuracy of estimates deducted from LIDAR metrics. A major one is the pulse density that affects directly the resolution of the returned signal, as well as the cost of the whole operation (Lim et al. 2008, Naesset 2009). Field position error could also affect estimates accuracy (Gobakken and Naesset 2009). As LIDAR metrics are calibrated using field positions generally measuredwith GPS, a positioning variability could affects results, especiallyin heterogeneous stands, where large variations may occurs over short distances.
Therefore, in this study we explored the use of multi-echo LIDAR metrics in estimating Ho forFrench deciduous stands of different structures. The impact of pulse densities was evaluated by thinning the original LIDAR dataset, and the effect of field position error was examined both indirectly by varyingthe LIDAR extraction area, and directly, on a subset of plots more precisely localised by triangulation from ground featureseasily observable on the LIDAR digital terrain models (DTM).
Material and Methods
Multi-echo LIDAR data were acquired at leaf-off state in March 2007, over 112 km2 of the Haye’s forest, in NorthEastern France (48°41’39’’N, 6°04’20’’E). LIDAR return densities ranged from 10 to 64 points/m2 which were classified as ground or vegetation by the vendor using a proprietary algorithm. From ground points, 3 digital terrain models (DTM) were built. One was provided by the vendor, and two others were modelled from a linear and a quadratic regression usingthese ground points. Flight characteristics are shown in Table 1.
LIDAR-based predictors were then generated by subtracting vegetation points from DTM. At plot level, height percentiles, returns densities and distribution parameters were calculated. Local maxima were also calculated using 2 methods. One was by subdividing the plot area into concentric rings of equal areas, and averaging the maximum heights obtained for each ring. This LIDAR metric is further identified with the prefix Hdoman_. The other method start by finding a maximal point and excludes an area of a given radius away from this point prior to find another maxima. This loops until the desired number of local maxima is obtained. Then, these heights are averaged. This metricis an attempt to avoid the bias of measuring more than once the same tree. Itmimicstree crowns, and the exclusion radius varies as a function of pulse height. The regression equation has been obtained from crown radius measurements of 800 beech trees (R2=0.54). This variable is further identified with the prefix Hmv_.
In order to appreciate the impact of pulse density, ground and vegetation returns were thinned up to 5% per plot. LIDAR metrics were recalculated and the thinning effect on the residual error was examined.
The impact of the LIDAR extraction area and GPS uncertaintieswere also examined, by varying the LIDAR data extraction radius from 10 to 20 m.
Table 1. LIDAR and flight characteristicsGround plots
As ground truth, 120 field plots (600 m2 each) were installed and classified according to stand structure (high forest, coppice with standards, or intermediate between these two types). Coppice with standards is a two-story management system where among regularly cut hornbeam trees (“coppice”), some oak and beech trees called “standards” are left to grow as larger size timber. For each plot, Ho was calculated by averaging top height measurements of the 5 biggest trees per plot (Duplat and Perrotte 1981). These estimatestop height of the 100 biggest tree per hectare, taking into account bias associated with reduced sampling area (Pierrat et al. 1995, Garcia 1998). Two opposed height measurements per tree were performed using a Vertex. For bent trees, corrections were performed to take into account the horizontal distance between the observer and the ground projection of tree apex. Plot position was recorded usinga Geoexplorer XT GPS. In order to examine positioning error, characteristic features on DTM were used to precisely localise 30 of these plots using laser telemeters and triangulation. Compared to GPS measurements, the averaged distance between these plot centre was 3 m (from 1 to 7 m).
Results and Discussion
Univariate models performed with LIDAR metrics to estimate Ho are presented in Table 1. Nine of these models have a good accuracy, with a residual error smaller than 1 m, which is probably closed to the measurement error from ground plots. Most of these models (7) use local maxima variables, which suggest that spatial information from LIDAR returns improves the accuracy of Ho estimates. Among the LIDAR metrics associated to the return distribution, the heights of the 95th percentile, followed by the 99th percentile have also low residual errors (Table1). The standard deviation of the height distribution of LIDAR returns yield a poor residual error (>1m).
Based on these results, 2 local maxima metrics (Hdoman5 and Hmv5) and the 95thheight percentile were retained to build 3multivariate models.The goal was to remove from the residuals of these models the stand structure effect, in order to use a single equation over all the forest types of this study. Inclusion of the standard deviation of returns height distribution (std) improved these models, and therefore was kept in all of them. However it did not remove the stand structure effect.As we presumed that the main difference between forest types was related to crown roughness, the interquartile range between the 99th and the 95thheight percentiles was included in the models. It successfully removed the structural effect for the models build with Hdoman5 and Hmv5, but not with the 95th height percentile. So, it was kept in the former 2 models only. Thus the residual error for these 3 final models ranged from 0.76 to 0.82m.
Table 2. Univariate results giving for each LIDAR metrics, the residual standard error (RMSE) and the determination coefficient (R2). Prefixes using hdoman_ and hmv_ are local maxima metrics, those using hlid_ are heightpercentiles. Std represents the height standard deviation of the vegetation returns.The impact of thinning up to 5% the LIDAR signal did not deteriorate the precision of Ho estimates. This result may be caused by the very high pulse density obtained in this experiment. Others have shown degradationin accuracy occurring only at very low signal density, much lower than the ones tested here (Gobakken and Naesset 2008, Lim et al. 2008).
The use of different DTM did not also affect drastically the residual error, suggesting that in our conditions (i.e. mostly flat terrain) the use of the DTM supplied by the provider, or the ones built from linear or quadratic regressions was of minor significance in terms of accuracy, probably well under the measurements error performed in the field.
In order to figure out the impact of uncertainties in plot positioning, we varied the LIDAR extraction area, based on GPS coordinates, as well as on coordinates corrected from ground triangulation using DTM’s characteristic features. Although these corrections were small (1-7m) it only slightly improved accuracy of the model based on a 13.8 m extraction radii, which also corresponds to the ground plot sampled area (Figure 1). Error remainsalso low, for a slightly larger extraction area (15m in radius). However, a clear degradation in precision is observed for extraction radii much larger (e.g.20 m) or much smaller (e.g.10 m) than the one used for ground sampling. Thus, it seems that the impact of positioning error on Ho accuracy is only of minor importance given the relatively precise GPS used in this experiment.
Figure 1. Impact of positioning uncertainties onthe residual standard error of Ho estimations (RMSE). 4 extraction radii were used, ranging from 10 to 20m. The impact of GPS and DTM positioning (Lidar) is also compared for each given extraction area. Ho was estimated using 3 models, 2 based on local maxima metrics (i.e. Hmv5 and Hdoman5) and a third based on the 95th height percentile of the returns distribution.
Conclusion
In this experiment, Ho was estimated with a residual error smaller then 1m. This accuracy is probably reflecting ground measurement error. Also, the use of local maxima metrics in the models, allowed removal of stand structure effects in the residues, which could givemore robustness for the next mapping stageof this project. Even though maximal height returns has been shown to be a highly variable LIDAR metric, affected for example by pulse density (Lim et al 2008, Naesset 2009), it appears in this study that local maxima metrics performed well in estimating Ho.Finally, uncertainties associated to GPS positioning were of minor significance in terms of Ho accuracy, probably due to the relatively precise instrument used.
Literature Cited
Duplat, P.andG. Perrotte. 1981. Inventaire et estimation de l’accroissement des peuplements forestiers. Office National des Forêts (ed.). Paris. 432 p.
Garcia, O. 1998. Estimating top height with variable plot size. Can. J. For. Res. 28: 1509-1517.
Gobakken, T. and E. Naesset. 2008. Assessing effects of laser point density, ground sampling intensity, and field sample plot size on biophysical stand properties derived from airborne laser scanner data. Can. J. For. Res. 38: 1095-1109.
Gobakken, T. and E. Naesset. 2009. Assessing effects of positioning errors and sample plot size on biophysical stand properties derived from airborne laser scanner data. Can. J. For. Res. 39: (in press).
Lim, K., C. Hopkinson and P. Treitz. 2008. Examining the effects of sampling point densities on laser canopy height and density metrics. Forestry Chronicle 84: 876-884.
Naesset, E. 2002. Predicting forest stand characteristics with airborne scanning laser using a practical two-stage procedure and field data. Remote Sensing of Environment 80: 88-99.
Naesset, E. 2009. Effects of different sensors, flying altitudes, and pulse repetition frequencies on forest canopy metrics and biophysical stand properties derived from small-footprint airborne laser data. Remote Sensing of Environment 113: 148-159.
Pierrat, J.-C., F. Houllier, J.-C. Hervé and R. Salas Gonzalez. 1995. Estimation de la moyenne des valeurs les plus élevées d’une population finie: application aux inventaires forestiers. Biometrics 51: 679-686.