Individual Tree Species Identification Using Lidar Intensity Data - asprs

More documents

Recommendations

Info

<strong>Intensity</strong> analysis between Broadleaved and Coniferous species To assess the separability of each variable for broadleaved and coniferous species, linear discriminant analysis was performed for leaf-on, leaf-off, and combined datasets. The results are shown in Table 5. Generally, variables computed using leaf-off data showed better classification accuracy than leaf-on data. Overall classification accuracy was improved by combining leaf-on and leaf-off datasets. Among variables computed using the leaf-on dataset, mean intensity values for the whole crown using all returns (whole_all) showed the best classification accuracy (71.2 %) while among variables computed using the leaf-off dataset, mean intensity values for the upper crown showed the best accuracy (80.3 %). When combining leaf-on and leaf-off datasets, mean intensity values for the whole crown using all returns showed the best classification accuracy (93.3%). To compare classification accuracy between leaf-on, leaf-off and combined datasets, appropriate variables were selected using principal component analysis (PCA). With derived variables, linear discriminant analysis (LDA) was performed for each dataset and the results are shown in Table 6. The accuracy using the leaf-off dataset was better than that using the leaf-on dataset. When combining leaf-on and leaf- off datasets, the accuracy was improved. The leaf-off dataset in this study contained species with flowers, Magnolia, Malus, and Prunus, and deciduous coniferous species, larch. After excluding these four species, the remaining deciduous broadleaved and evergreen coniferous species were tested for classification accuracy using LDA for each dataset (see Table 6). For leaf-off data, the sorted species improved the classification accuracy up to 96.9 % and for the combined datasets, up to 98.2 %. The classification accuracy decreased for leaf-on data using these sorted species. Table 5. Classification accuracy for broadleaved and coniferous species for each variable using linear discriminant analysis with leaf-on, leaf-off and combined datasets. Classification accuracy (%) Variables Leaf-on dataset Leaf-off dataset All datasets whole_all 71.2 78.5 93.3 whole_1 69.9 79.4 91.0 upper_all 61.3 80.3 89.7 upper_1 60.9 80.3 90.1 surface_all 66.3 79.4 87.0 surface_1 68.1 79.8 87.0 cv_all 56.2 69.0 75.4 cv_1 69.9 71.7 82.0 prop_1 54.2 66.9 69.8 Table 6. Classification accuracy for broadleaved and coniferous species with all species and sorted species using linear discriminant analysis with leaf-on, leaf-off and combined datasets. Classification accuracy (%) Variables Leaf-on dataset Leaf-off dataset All datasets All species 68.6 82.5 88.8 Sorted species 57.8 96.9 98.2 DISCUSSION Broadleaved species showed higher mean intensity values than coniferous species in leaf-on data. Song et al. (2002) reported the same result using LIDAR intensity data. Spectral reflectance of broadleaved species was found to be higher than that of coniferous species in several studies in the near-infrared wavelength region. Roberts et al. (2004) found that spectral reflectance of five broadleaved deciduous species was higher than five conifers in the near-infrared wavelength region at the branch-scale. A recent finding that intensity data were directly related to spectral reflectance of the target materials (Ahokas et al., 2006) implies that the result of LIDAR intensity analysis in this research is consistent with the results of spectral reflectance studies. Different intensity values between broadleaved species and coniferous species are probably related not only to the reflective properties of the vegetation, but also to the larger scale properties of the forest such as canopy openness and the spacing and type of ASPRS 2008 Annual Conference Portland, Oregon April 28 - May 2, 2008
foliage components within individual tree crowns. For example, laser pulses probably have a greater chance of passing through needles or scale-like leaves than broad leaves. A pulse passing deeper into the crown would generate multiple returns for each echo and consequently mean intensity values would be lower within the crown and more likely associated with branches. In contrast, broadleaved species showed lower mean intensity values than coniferous species in leaf-off data. This result is supported by Roberts et al. (2004) who reported that bark has a lower spectral reflectance than leaves. In leaf-off conditions, most laser pulses would reflect from woody materials such as stems, branches and bark lowering reflectance of deciduous species relative to non-deciduous species. In the leaf-on data analysis, Betula showed very low intensity values among broadleaved species probably because Betula has relatively sparse foliage increasing the chance that laser pulses reflect from branches. In contrast, Prunus showed the highest mean intensity values in leaf-on data probably because foliage is distributed densely within the crown, laser pulses would most likely reflect from the crown surface which is mostly composed of leaves rather than branches, resulting in higher intensity values. Among coniferous species, Douglas-fir, western hemlock, redwood and spruce showed higher intensity values than larch, cedar and Pinus in both leaf-on and leaf-off data. This result seems to be related to their leaf structures. The former four species have single needles, whereas Pinus and larch have clustered needles and cedar has small scale-like needles. <strong>Species</strong> with single needles showed higher intensity than those with clustered needles probably because clustered needles allow bare branches to be exposed between needle clusters, which would raise the chance of laser pulses reflecting from branches. Overall, pairs of tree species showed very significant differences for the mean intensity values for the whole crown in leaf-off data, even within broadleaved species and within coniferous species. Foliar conditions of evergreen coniferous species are not very different between two datasets, however, the pair-wise significance tests among coniferous species in leaf-on data showed poorer separability than in leaf-off data. This is probably because two LIDAR datasets were acquired using different laser scanners operated by different vendors. The laser scanner system used to acquire leaf-off data, Optech ALTM 3100, with a higher pulse repetition frequency, higher point density per square meter and more overlap between flight lines, could differentiate each species better than that used to acquire leaf-on data, Optech ALTM 30/70. The result of discriminant analysis implies that combining leaf-on and leaf-off data would result in better overall classification accuracies for broadleaved species and coniferous species than using data from one season. However, trade-offs should be considered between the cost of acquiring additional LIDAR data and the improvement in the classification accuracy. In fact, combining leaf-on and leaf-off data didn’t improve classification accuracy substantially compared with using the leaf-off data only. Therefore, for the purpose of classifying these two species groups, using LIDAR data acquired in leaf-off conditions is recommended. The LIDAR dataset obtained in mid-March didn’t represent ideal leaf-off conditions for some species. Predictably, classification accuracy was improved by deleting three broadleaved species that were flowering at the time of the acquisition, Magnolia, Malus and Prunus, and one deciduous coniferous species, larch, using the leaf-off data. This result implies that if LIDAR data could be acquired during the winter with complete leaf-off conditions, the distinction between leaf-on and leaf-off conditions would be more reliable, resulting in better classification accuracy for broadleaved and coniferous species. Or, because trees with leaves or those without leaves could be recognized via aerial photographs and photos taken at the field at the time of the LIDAR data acquisition on March 17th, if we restricted our investigation to these clearly distinguished species, with leaves or without leaves, the analysis using leaf-on and leaf-off conditions would be more reliable. The classification accuracy decreased without these four species in the leaf-on dataset. This is probably because these species are not different from the others in late summer and only reduced the overall sample size. The presence or absence of foliage in deciduous species changes seasonally and the time of blooming varies depending on species. It would be helpful to know when each species blooms prior to selecting the date for data acquisition. CONCLUSIONS AND FUTURE WORK The overall results showed that LIDAR intensity data could distinguish broadleaved species from conifers and further distinguish various tree species within these broad groups. Different intensity values between species were related not only to reflective properties of the vegetation, but also to a presence or absence of foliage and the arrangement of foliage and branches within individual tree crowns. Two different seasonal LIDAR datasets resulted in different relative intensity values among species with better separation using leaf-off data than leaf-on data, albeit with two different LIDAR sensors. Future directions for this study will include acquiring LIDAR data in both summer and winter, using the same LIDAR system to control for system effects. This will lead to more reliable ASPRS 2008 Annual Conference Portland, Oregon April 28 - May 2, 2008
Page 1 and 2: INDIVIDUAL TREE SPECIES IDENTIFICAT
Page 3 and 4: Figure 1. Approximate location of t
Page 5 and 6: Table 2. Summary of field measureme
Page 7 and 8: Principal Component Analysis. Princ
Page 9: (a) (b) Table 3. The result of t-st

Individual Tree Species Identification Using Lidar Intensity Data - asprs

Create successful ePaper yourself

Delete template?

Save as template?