Can global navigation satellite system signals reveal the ecological attributes of forests?

International Journal of Applied Earth Observation and Geoinformation 50 (2016) 74–79

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation and Geoinformation journal homepage: www.elsevier.com/locate/jag

Can global navigation satellite system signals reveal the ecological attributes of forests? Jingbin Liu a,b,∗∗ , Juha Hyyppä a,b , Xiaowei Yu a,b , Anttoni Jaakkola a,b , Xinlian Liang a,b,∗ , Harri Kaartinen a,b , Antero Kukko a,b , Lingli Zhu a,b , Yunsheng Wang a,b , Hannu Hyyppä b,c a

Department of Remote Sensing and Photogrammetry, Finnish Geospatial Research Institute, 02430 Masala, Finland Center of Excellence in Laser Scanning Research, Finnish Geospatial Research Institute, 02430 Masala, Finland c Research Institute of Modelling and Measuring for the Built Environment, Aalto University, Espoo 00076, Finland b

a r t i c l e

i n f o

Article history: Received 2 December 2015 Received in revised form 17 March 2016 Accepted 18 March 2016 Available online 24 March 2016 Keywords: Global navigation satellite system Ecological attributes Crowdsourcing Forest inventory Laser scanning Remote sensing

a b s t r a c t Forests have important impacts on the global carbon cycle and climate, and they are also related to a wide range of industrial sectors. Currently, one of the biggest challenges in forestry research is effectively and accurately measuring and monitoring forest variables, as the exploitation potential of forest inventory products largely depends on the accuracy of estimates and on the cost of data collection. A low-cost crowdsourcing solution is needed for forest inventory to collect forest variables. Here, we propose global navigation satellite system (GNSS) signals as a novel type of observables for predicting forest attributes and show the feasibility of utilizing GNSS signals for estimating important attributes of forest plots, including mean tree height, mean diameter at breast height, basal area, stem volume and tree biomass. The prediction accuracies of the proposed technique were better in boreal forest conditions than those of the conventional techniques of 2D remote sensing. More importantly, this technique provides a novel, costeffective way of collecting large-scale forest measurements in the crowdsourcing context. This technique can be applied by, for example, harvesters or persons hiking or working in forests because GNSS devices are widely used, and the field operation of this technique is simple and does not require professional forestry skills. © 2016 Elsevier B.V. All rights reserved.

1. Introduction The total global forest area exceeds 4 billion hectares, and it covers 31% of the total land surface (FAO, 2010). Additionally, forests contain more than half of all terrestrial species and account for 75% of terrestrial gross primary production (GPP), which represents carbon assimilation from photosynthesis by vegetation per unit area and time (Beer et al., 2010; Pan et al., 2013), and 80% of Earth’s total plant biomass (Kindermann et al., 2008). All of these reflect the high ecological and economic importance of forests. The conversion of solar energy into biomass through photosynthesis makes forest ecosystems a key component of the global carbon cycle and climate. A major portion of the total forest carbon storage comprises the growing stock of carbon reserves (Kindermann et al., 2008). Thus, one of the biggest current challenges in forest inventory research is effectively and accurately measuring and

∗ Corresponding author at: Department of Remote Sensing and Photogrammetry, Finnish Geospatial Research Institute, 02430 Masala, Finland. ∗∗ Corresponding author. E-mail address: xinlian.liang@nls.fi (X. Liang). http://dx.doi.org/10.1016/j.jag.2016.03.007 0303-2434/© 2016 Elsevier B.V. All rights reserved.

monitoring forest biomass (Holopainen et al., 2014; Hyyppä et al., 2008; Liang et al., 2014). The recent knowledge regarding forest biomass is based on ground measurements and coarse- or mediumresolution satellite images (Hyyppä et al., 2008; Liang et al., 2014; Thenkabail 2015). The exploitation potential in the field of biomass and stem volume mapping is largely dependent on the accuracy of what can be obtained and the cost of data collection. Optimally, when forest inventory data are more accurate and updated in time, the storages of the forest industry will move from real storages to forests (Holopainen et al., 2014; Liang et al., 2014). There is also growing interest for citizens to participate in the data collection of geospatial information (Heipke, 2010). More commonly, crowdsourcing is understood as geospatial data collection by voluntary citizens who are untrained in the disciplines of geography, cartography or related fields (Thenkabail, 2015). In the field of forestry, crowdsourcing has been used to assess the condition of city trees. For example, PhillyTreeMap is a web-based application that allows citizens to input tree information of city forests; today, data on over 179,000 Philadelphian trees are stored in the database. A variety of field measurement and remote sensing technologies have been studied in the prediction of forest biophysical attributes and their changes (Koch, 2010; Rahlf et al., 2014; Hyyppä et al.,

J. Liu et al. / International Journal of Applied Earth Observation and Geoinformation 50 (2016) 74–79 Table 1 Correlations (R) between GNSS signal strength losses and forest attributes. Forest attributes

GPS

GLONASS

Mean tree height Mean DBH Basal area Stem volume Biomass

0.75 0.73 0.85 0.84 0.85

0.71 0.69 0.78 0.76 0.77

2008; Liang et al., 2012, 2014, 2015; Karila et al., 2015; Solberg et al., 2015; Nilsson, 1996; Yu et al., 2010; Nelson et al., 1988). Reference data for sample plots are conventionally collected through manual measurements, which are expensive, time consuming and labor intensive (Holopainen et al., 2014; Heipke, 2010). Consequently, trees are typically measured by sampling criteria already at the plot-level, and tree attributes are often retrieved using allometric models instead of actual measurements, which introduce inaccuracies that propagate to the stand- and national-scale estimation of forest resources(Holopainen et al., 2014; Karila et al., 2015; Solberg et al., 2015). Previous studies also show that GNSS (global navigation satellite system) positioning does not work well in forests (Sigrist et al., 1999; Kaartinen et al., 2015; Alonso et al., 2014). The denser the canopy is, the less accuracy the positioning solution possesses (Meng et al., 2009; COST235, 1996). Therefore, it is surprising that GNSS signals have not yet been studied as an observable source to measure forest attributes. In this study, we asked, “Can GNSS signals be used as an informative indicator to reveal the physical properties of forests?” This is not only a scientific question; it also may have remarkable societal and industrial impacts.

75

Table 2 Accuracy of plot-based forest attribute predictions using the features of different GNSS combinations. Features

Forest attributes

GPS

0.00 0.00 Mean height (m) Mean DBH (cm) −0.06 −0.25 BA (m2 /ha) −0.01 −0.07 0.70 0.32 Volume (m3 /ha) Biomass (Mg/ha) −0.37 −0.37

Bias

GLONASS

Mean height (m) Mean DBH (cm) BA (m2 /ha) Volume (m3 /ha) Biomass (Mg/ha)

−0.10 −0.11 −0.12 −1.38 −0.44

GPS + GLONASS

Mean height (m) Mean DBH (cm) BA (m2 /ha) Volume (m3 /ha) Biomass (Mg/ha)

−0.08 −0.05 −0.08 −0.39 −0.65

Bias (%)

RMSE

RMSE (%)

R

4.05 4.25 6.03 77.94 34.04

18.77 19.54 30.78 35.81 33.51

0.83 0.79 0.83 0.82 0.83

−0.47 −0.51 −0.64 −0.63 −0.43

4.34 4.74 6.56 89.40 39.07

19.92 21.49 33.53 40.79 38.38

0.80 0.74 0.79 0.75 0.77

−0.38 −0.22 −0.44 −0.18 −0.66

3.61 4.13 5.57 72.78 32.14

16.95 19.24 29.22 34.39 32.53

0.87 0.80 0.86 0.85 0.85

test site, including ranging measurements, navigation data, and signal strength indicators. One was placed on top of a tripod under open-sky conditions (Supplementary Fig. S2), and it was the benchmark for calculating the signal strength loss (SSL), which is the primary observable of estimating forest attributes in this study; the other, called the rover receiver, was carried on an all-terrain vehicle (Supplementary Fig. S3), which was driven around the test site for collecting GNSS data. Both GNSS receivers were able to track the US Navstar GPS (Global Positioning System) and the Russian GLONASS (GLObalnaja NAvigatsionnaja Sputnikovaja Sistema) satellites. 3. Derivation of GNSS features

2. Material and methods 2.1. Test site and reference dataset of forest inventory The proposed technique was studied using field experiments conducted at our established test site located in Evo, southern Finland (61.19◦ N, 25.11◦ E), where a reference dataset of forest inventory has been established. The test site is part of the southern Boreal Forest Zone, and it mainly comprises approximately 2000 ha of managed boreal forest. The topography of the area varies from 125 m to 185 m above sea level. Scots pine (Pinus sylvestris) and Norway spruce (Picea abies) are the dominant tree species in the study area, contributing 40% and 35% of the total volume, respectively (Yu et al., 2010). The reference dataset of forest attributes included mean tree height, mean diameter at breast height (DBH), basal area, stem volume, and biomass of trees. These data were generated with airborne laser-scanning data (acquisition date July 25th 2009, point density 10 points per m2 ) and field measurements in 292 sample plots (10 m radius circular area distributed over the same area as the study area) using an area-based approach (Hyyppä et al., 2008; Yu et al., 2010) and Random forests (RF) technique (Breiman, 2001; Hastie et al., 2009; Yu et al., 2011). Finally, plots of circular areas of 10 m radius were created along the trajectory with the center point of the circle following the trajectory, and they were utilized as reference plots. In this study, the reference datasets of these plots were used as sample set in the RF regression algorithm to establish the proposed GNSS-based nonparametric regression model for predicting forest attributes. 2.2. GNSS data collection As illustrated by Supplementary Fig. S1, two Trimble GNSS receivers were used to collect GNSS observables in parallel in the

The distributions of the received signal powers of both receivers were compared in terms of carrier-to-noise ratio (C/N0 ) as shown in Supplementary Fig. S4 and evidently showed lower signal strengths with the in-forest receiver overall. Due to the presence of forest canopies, the signal strengths received by the in-forest receiver were less stable than those of the out-forest one, as shown in Supplementary Fig.S5, which compares the variations between two adjacent epochs of the observed signal strengths of both receivers. It shows that the in-forest receiver at 9.4% epochs was subject to C/N0 fluctuations greater than 3 dB, while this value was 1.3% epochs for the out-forest receiver. The out-forest receiver was utilized as the benchmark of the signals in space, and the signal strength loss was defined by the difference between the two receivers in C/N0 observables at each corresponding epoch (t) for each satellite (sv), i.e.,SSL(sv,t) = [C/N 0in−forest − C/N0out−forest ](sv,t) . SSL measurements were calculated for every satellite and epoch, and the satellites of low elevations were excluded by a cut-off elevation of 30◦ because they have higher noise levels in the measurements (Wu et al., 2010). As a function of satellite elevations, the slant penetration paths of the satellite signals caused larger signal strength losses than the vertical path. A mapping function was utilized for compensating the effect of satellite elevations and normalizing SSL measurements to the vertical direction, SSL⊥ = SSL × sin(), where  is the satellite elevation and SSL⊥ is the adjusted signal strength loss. Every SSL measurement was associated with a specific sample plot of forest reference data to which the rover receiver’s position of current epoch belonged. Thus, there were a number of SSL measurements within the range of each sample plot. For each sample plot and each GNSS system (GPS and GLONASS in this study), six features of SSL measurements were derived, as indicated below, for the prediction of forest attributes: • number of SSL measurements

76

J. Liu et al. / International Journal of Applied Earth Observation and Geoinformation 50 (2016) 74–79

Fig. 1. Scatterplots of predicted versus reference values for plot-level forest attributes: (a) mean tree height, (b) mean DBH, (c) basal area, (d) stem volume, and (e) tree biomass, derived with the features of GPS, GLONASS, and combined GPS and GLONASS (labelled with “GNSS”), respectively. The line indicates a 1:1 relationship.

• • • • •

arithmetic mean value of SSL measurements standard deviation (STD) of SSL measurements median value of SSL measurements maximum value of SSL measurements minimum value of SSL measurements

4. Feasibility and numerical results of predicting forest attributes using GNSS signals Table 1 shows the correlation coefficients (R) between the mean values of the SSL measurements and five important attributes for over 300 sample plots where the SSL measurements were collected.

J. Liu et al. / International Journal of Applied Earth Observation and Geoinformation 50 (2016) 74–79

77

Fig. 2. Statistical probability densities of prediction errors of five plot-level forest attributes: (a) mean tree height, (b) mean DBH, (c) basal area, (d) stem volume, and (e) tree biomass, derived with the features of GPS, GLONASS, and combined GPS and GLONASS (labelled with “GNSS”), respectively.

The high correlation coefficients ranging from 0.7–0.85 for both constellations imply that the GNSS signal strength loss can be an effective indicator for revealing forest properties. The correlations were more significant for the basal area, stem volume and biomass than for the mean tree height and DBH, indicating the GNSS signaling can provide complementary information to 3D techniques which are all based on tree height (Rahlf et al., 2014; Yu et al., 2015). Using the six features of GNSS SSL listed above, the Random forests nonparametric regression technique is applied to the prediction of forest attributes (Breiman 2001; Yu et al., 2011). Supplementary Tables S1–S3 list the features defined with the GPS, GLONASS, and combined GPS and GLONASS datasets. The estimates

of plot-level forest attributes with GNSS features were compared with the established reference dataset. Fig. 1 shows the scatterplots of the predicted versus reference values for the mean tree height, mean DBH, basal area, stem volume and biomass of the sample plots using GPS, GLONASS, and the combination of GPS and GLONASS features. The accuracies of the predicted forest attributes using different combinations of GNSS features were presented in Table 2, and the probability distributions of prediction errors of these attributes were shown in Fig. 2. The mean tree height was estimated with a relative root-meansquare error (RMSE) of 18.77% and an R-value of 0.83 using the GPS features and a relative RMSE of 19.92% and an R-value of 0.8

78

J. Liu et al. / International Journal of Applied Earth Observation and Geoinformation 50 (2016) 74–79

using the GLONASS features. When the features of both GNSS systems were combined, the corresponding values were 16.95% and 0.87, respectively. In Fig. 1(a), biases were observed for plots with different ranges of tree heights. The predictions were smaller than the reference values for most of plots with mean tree heights of lower than 15 m or higher than 20 m, whereas the predictions were greater than the references for most of plots with mean tree heights between 15 and 20 m. This was partially caused by biases in the reference data of mean tree height. This study generated the reference data using airborne laser scanning data, and it was a known issue that mean tree heights may be underestimated or overestimated for different situations due to errors of tree segmentation algorithms (Yu et al., 2011). Predictions of DBH, with GPS features, had a relative RMSE of 19.54% and an R-value of 0.79, whereas they were 21.49% and 0.74 using GLONASS features. The combination of GPS and GLONASS features slightly improved the prediction of DBH with an RMSE of 19.24% and an R-value of 0.8. The basal area predicted using GPS features had approximately 3% smaller RMSEs and 0.04 higher R values than those using GLONASS features, and the combined features of both GNSS systems slightly improved the estimations with 1.5% smaller RMSEs and 0.03 higher R values than the GPS-derived results. For stem volume and biomass, the predictions with GPS features had approximately 5% smaller relative RMSEs and 0.06 higher R values than the results derived with GLONASS features, and the combination of GPS and GLONASS features slightly improved the estimations with 1–1.5% smaller relative RMSEs and 0.02 higher R values than the GPS-derived results. The biases and relative bias ratios between the predicted values and the reference dataset were less than ±0.7% for all forest attributes, as shown in Table 2. For specific plots, errors in the predictions of forest attributes were mainly caused by noisy measurements of signal strength losses, which were indicated by noisy fluctuations of signal strength observables in Fig. S5. Statistical probability densities of the prediction errors followed approximately symmetric and bell-shaped probability distribution curves with a single peak, which were like normal distributions, as showed in Fig. 2. Prediction errors of all these attributes have mean values close to zero, i.e. the bias values in Table 2. The prediction errors have the most probabilities around the bias values, and the probability densities quickly decrease as the magnitudes of prediction errors deviate away from the zero values. Conventional remote sensing techniques, such as using image intensities, textures and NDVI (Normalized Difference Vegetation Index)-type features, were compared in Hyyppä et al. (2000), who concluded that 2D remote sensing technologies have a reasonable, modest explanatory power to assess forest characteristics. Comparing the GNSS derived results of basal area with those obtained in Hyyppä et al. (2000), the GNSS-based predictions were as accurate as the best 2D remote sensing data sources, such as imaging spectrometry and aerial imagery, and far more accurate than satellite-based 2D techniques. When comparing the results with those of the 3D techniques in Yu et al. (2015), who extracted canopy height information from airborne images, laser scanning and various satellite technologies, the first impression is that 3D techniques are more accurate. However, all 3D techniques produce a notable cost to the forest inventory.

and tree biomass. Compared with the reference dataset of forest inventory in an established test site, the accuracies of the predicted forest attributes with GNSS signals were evaluated with bias errors, RMSEs and Pearson’s correlations between the predicted and reference values. The results show that when using all GNSS features, the predicted values of the mean tree heights of the sample plots had a relative RMSE of 16.95% and a correlation (R) value of 0.87, and the corresponding values were 19.24% and 0.8 for mean DBH. For the predictions of plot-level basal area, stem volume and biomass, RMSE values were 29.22%, 34.39%, and 32.53%, respectively, and the correlations between predicted and reference values were approximately 0.85. The bias errors are less than ±0.7% for all forest attributes. Overall, the combination of GPS and GLONASS signals generated better accuracies than any single system for all forest attributes. It is reasonable to expect that the predictions will be improved by combining more emerging constellations of global and regional satellite navigation systems when they become operational in the future, e.g., the European Galileo and Chinese Beidou systems coming into full operational capability by 2020. The results show that the estimates of forest attributes using GNSS-based features are comparable to the best 2D professional remote sensing technologies. The finding that GNSS can be used to measure important stand attributes requires further usability tests. GNSS devices are largely available. There are ¾ million forest owners alone in Finland (14% of the population) (Tokola et al., 2007). Many of them visit their forests regularly or live close to their forests; thus, personal technologies could be used to collect reference information of their own forests. In addition, a large number of people hike in boreal forests, and there are 1000 harvesting companies alone in Finland (Tokola et al., 2007), where GNSS information is collected constantly. The operation of GNSS data collection is as simple as ordinary positioning, and it does not require extra professional skills. Additionally, improved GNSS features can presumably be found to obtain improved accuracy. Alternatively, other publicly available data sources, e.g., the national open airborne laser scanning (ALS) database and digital surface models (DSM), can be utilized for extracting tree height features and integration with GNSS signals. Therefore, the proposed technique is a potential low-cost solution for generating large-area reference data of forest inventories in the future in a crowdsourcing context.

Author contributions J.L. designed the experiment, processed the GNSS data, and analyzed the results. J.H. and A.J. proposed the scientific concept and J.H. wrote part of the paper. X.Y. calculated the prediction using the RF technique. A.J. and X.L. contributed to the scientific content and data-processing techniques. All co-authors contributed to the field experiments and data processing and analysis. J.L. led the writing of the paper with substantial input from all co-authors.

Competing financial interests The authors declare no competing financial interests.

Acknowledgements 5. Conclusions and discussions This communication proposes GNSS signals as a novel type of observable data for predicting forest attributes. It analyses the feasibility by showing high correlations in boreal forest conditions between forest attributes and the variables of GNSS signals and presents the prediction results of plot-level forest attributes, including mean tree height, mean DBH, basal area, stem volume

This work was mainly funded by Academy of Finland project “Interaction of Lidar/Radar Beams with Forests Using Mini-UAV and Mobile Forest Tomography” (No. 259348). Financial support of some of the co-authors was also obtained from Academy of Finland projects “Centre of Excellence in Laser Scanning Research (CoELaSR)” (No. 272195) and “Competence-Based Growth Through Integrated Disruptive Technologies of 3D Digitalization, Robotics,

J. Liu et al. / International Journal of Applied Earth Observation and Geoinformation 50 (2016) 74–79

Geospatial Information and Image Processing/Computing − Point Cloud Ecosystem” (No. 293389). Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.jag.2016.03.007. References Alonso, A.A., et al., 2014. Dual-polarization, GNS interference pattern technique for soil moisture mapping. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 7 (5), 1533–1544, http://dx.doi.org/10.1109/JSTARS.2014.2320792. Beer, C., et al., 2010. Terrestrial gross carbon dioxide uptake: global distribution and covariation with climate. Science 329, 834–838. Breiman, L., 2001. Random forests. Mach. Learn. 45, 5–32. COST235, (1996). Radio propagation effects on next-generation fixed-service terrestrial telecommunication systems, Final Report, Luxembourg. FAO, 2010. Global forest resources assessment 2010: main report. In: FAO Forestry Paper 163. Food and Agriculture Organization of the United Nations, Rome. Hastie, T., et al., 2009. The Elements of Statistical Learning: Data Mining, Inference, and Prediction, 2nd edition). Springer-Verlag, New York. Heipke, C., 2010. Crowdsourcing geospatial data. ISPRS J. Photogramm. Remote Sens. 65 (6), 550–557. Holopainen, M., et al., 2014. Outlook for the next generation’s precision forestry in Finland. Forests 5, 1682–1694. Hyyppä, J., et al., 2000. Accuracy comparison of various remote sensing data sources in the retrieval of forest stand attributes. For. Ecol. Manag. 128 (1), 109–120. Hyyppä, J., et al., 2008. Review of methods of small-footprint airborne laser scanning for extracting forest inventory data in boreal forests. Int. J. Remote Sens. 29 (5), 1339–1366. Kaartinen, H., et al., 2015. Accuracy of kinematic positioning using global satellite navigation systems under forest canopies. Forests 6, 3218–3236. Karila, K., et al., 2015. Tandem-X interferometry in the prediction of forest inventory attributes in managed boreal forests. Remote Sens. Environ. 159, 259–268. Kindermann, G.E., et al., 2008. A global forest growing Stock, biomass and carbon map based on FAO statistics. Silva Fennica 42, 387–396. Koch, B., 2010. Status and future of laser scanning, synthetic aperture radar and hyperspectral remote sensing data for forest biomass assessment. ISPRS J. Photogramm. Remote Sens. 65 (6), 581–590.

79

Liang, X., et al., 2012. Detecting changes in forest structure over time with Bi-temporal terrestrial laser scanning data. ISPRS Int. J. Geo-Inf. 1 (3), 242–255, http://dx.doi.org/10.3390/ijgi1030242. Liang, X., et al., 2014. The use of a mobile laser scanning system for mapping large forest, plots. I.E.E.E. Geosci. Remote Sens. Lett. 11, 1504–1508, http://dx.doi. org/10.1109/LGRS.2013.2297418. Liang, X., et al., 2015. Forest data collection using terrestrial image-based point clouds from a handheld camera compared to terrestrial and personal laser scanning. IEEE Trans. Geosci. Remote Sens. 53, 5117–5132, http://dx.doi.org/ 10.1109/TGRS.2015.2417316. Meng, Y.S., et al., 2009. Empirical near ground path loss modeling in a forest at VHF and UHF bands. IEEE Trans. Antennas Propag. 57 (5), 1461–1468, http://dx.doi. org/10.1109/TAP.2009.2016703. Nelson, R., et al., 1988. Estimating forest biomass and volume using airborne laser data. Remote Sens. Environ. 24, 247–267. Nilsson, M., 1996. Estimation of tree heights and stand volume using an airborne Lidar system. Remote Sens. Environ. 56 (1), 1–7. Pan, Y., et al., 2013. The structure, distribution, and biomass of the world’s forests. Annu. Rev. Ecol. Evol. Syst. 44, 593–622. Rahlf, J., et al., 2014. Comparison of four types of 3D data for timber volume estimation. Remote Sens. Environ. 155, 325–333. Sigrist, P., et al., 1999. Impact of forest canopy on quality and accuracy of GPS measurements. Int. J. Remote Sens. 20, 3595–3610. Solberg, S., et al., 2015. Estimating forest biomass from TerraSAR-X stripmap radargrammetry. IEEE Trans. Geosci. Remote Sens. 53 (1), 154–161. Thenkabail, P.S., 2015. Land resources monitoring, modeling, and mapping with remote sensing. In: in a Series of Remote Sensing Handbook. CRC Press (ISBN 9781482217957). Tokola, T., et al., 2007. Forest reconnaissance surveys: comparison of estimates based on simulated TerraSar and optical data. Photogramm. J. Finl. 20 (2), 64–79. Wu, Y., et al., 2010. Cycle slip detection using multi-frequency GPS carrier phase observations: a simulation study. Adv. Space Res. 46 (2), 144–149. Yu, X., et al., 2010. Comparison of area-based and individual tree-based methods for predicting plot-level forest attributes. Remote Sens. 2 (6), 1481–1495, http://dx.doi.org/10.3390/rs2061481. Yu, X., et al., 2011. Predicting individual tree attributes from airborne laser point clouds based on random forests technique. ISPRS J. Photogramm Remote Sens. 66, 28–37. Yu, X., et al., 2015. Comparison of laser and stereo optical, SAR and InSAR point clouds from air- and space-Borne sources in the retrieval of forest inventory attributes. Remote sens. 7 (12), 15933–15954, http://dx.doi.org/10.3390/ rs71215809.