1. SINGLE-TREE FOREST INVENTORY USING LIDAR AND AERIAL IMAGES FOR 3D TREETOP POSITIONING, SPECIES RECOGNITION, HEIGHT AND CROWN WIDTH ESTIMATION Ilkka Korpela University of Helsinki I. Korpela, B. Dahlin, H. Schäfer, E. Bruun, F. Haapaniemi, J. Honkasalo, S. Ilvesniemi, V. Kuutti, M. Linkosalmi, J. Mustonen, M. Salo, O. Suomi, H. Virtanen

2. Contents Study objectives Methods – the STRS system Experimental STRS-based forest Inventory: Data, Field work, Results Conclusions

3. Study objectives Method development and testing in Single-Tree Remote Sensing, STRS Rationales: Forest inventory, 3D models, Landscape planning STRS-tasks: - 2D/3D positioning - Species recognition - Height estimation - Crown dimensions  Stem diameter, volume, bucking

4. Single-Tree Remote Sensing, STRS - Revised Airborne, active and passive instruments, 2D or 3D Direct estimation + indirect an allometric estimation phase Restrictions: tree discernibility, detectable object size, occlusion and shading, interlaced crowns → ILL-POSED Alternative or complement to field methods and area-based methods Accuracy restricted by “allometric noise” → tree and stand- level allometric bias, and tree-level imprecision, dbh ~ 10%. Measurements are subject to bias and imprecision Timber quality remains unsolved, only quantity Unsolved issues: Species recognition?

5. Methods – the STRS system we used Solves all of the STRS tasks Semi-automatic solutions combined with “operator intervention” Optional input: 1) images + DTM 2) images + LiDAR + DTM Assumptions: large-scale images, semi-dense LiDAR, accurate DTM Combine use of images and LiDAR for optimal results Constrain and filter using allometric regularities

6. Methods 1: Multi-scale template matching in 3D treetop positioning Assume that the optical properties and the shape of trees are invariant to their size.

7. Maxima at different scales, take global → (X,Y,Z) Treetop

8. Methods 2: Multi-scale Template matching – Crown width estimation in images Near-nadir views have been found best for the manual measurement of crown width in aerial images

9. Methods 3: Species recognition Spectral values Texture Variation in - Phenology - Tree age and vigor - Image-object-sun geometry => reliable automation problematic => bottleneck in STRS

10. Methods 4: LS-adjustment of a crown model with LiDAR points Assume that 1) Photogrammetric 3D treetop position is accurate 2) Trees have no slant 3) Crowns are ± rotation symmetric 4) We know tree height and species  approximation of crown size and shape → LiDAR hits are “observations of crown radius at a certain height below the apex” Assume a rather large crown and collect LiDAR hits in the vicinity of the 3D treetop position. Use LS-adjustment to find a crown model.

11. LiDAR hits are observations of crown radius at a certain height below the apex?

12. Example - a 22-m high birch: Solution in six iterations. Final RMSE 0.47 m RMSEs have been larger for birch in comparison to pine and spruce.

13. Methods 5: Allometric estimation of stem diameter and saw/pulp log volumes – stem bucking Species-specific regression equations that map maximal crown width (dcrm) and tree height (h) into dbh (Kallio virta and Tokola, 2005) Species-specific polynomial regression equations that model the tapering of stem diameter (Laasasenaho 1982)

14. Experimental forest inventory – STRS measurements 56.8-ha forest. 25-70 and 100-130- yr-old. CIR and PAN-RGBIR images with 12 cm and 9 cm GSD ALTM3100 LiDAR with 6-9 pulses per m2 A LiDAR-based DTM 5294 STRS-trees in 59 0.04-ha plots. 165 trees/hour with variables - Treetop position in XYZ - h, photogrammetric - h, LiDAR-based - Sp, photogrammetric - dcrm, photogrammetric - dcrm LiDAR-based

15. Experimental forest Inventory – FIELD measurements Before going to the field, process the STRS measurements into maps and tree labels

16. Experimental forest Inventory – FIELD measurements 1. Find the plot center using satellite positioning and the tree map 2. Find and label the STRS-trees using triangulation with a compass (solve commission errors) 3. Label the unseen trees in the circular 0.04-ha plot (you are left with omission errors) 4. Position in XY the unseen trees with trilateration and triangulation (ref. Silva Fennica 3/07), ~ 0.3 m. 5. Measure the trees for reference values of Sp, dbh and height

17. RESULTS – Which trees could we observe and position? - Omission errors: 38.8% in stem number (dbh > 50 mm) - Commission errors ~ 1-2% - Dominating trees were measurable, but not those with fused crowns - Less than 50% of the trees with a relative height of below 0.7 were detectable ”Tree discernibility” = proportion (%) of detected trees as a function of relative tree height.

18. Overall Sp-recognition accuracy was ~ 95%, which is at an acceptable level. RESULTS – Visual tree species recognition

19. - Multi-scale TM + LiDAR DTM ~ 0.71 m RMSE, 14-cm underestimation - Highest LiDAR hits + LiDAR DTM ~ 0.82 m RMSE, 58-cm underestimation Accuracy varied between species. DTM-errors were meaningless. RESULTS – Height estimation accuracy Residual plot

20. RESULTS – DBH estimation accuracy - Photogrammetric height, species and dcrm  28.7% RMSE with a 20-% underestimation - Photogrammetric height, species and LiDAR-based dcrm  19.6% RMSE with a 6-% underestimation Tests were not performed, where the dcrm- measurements had not been used.

21. RESULTS – Volume estimation accuracy With the use of images alone, an RMSE of 60% was observed for the single-tree stem volume estimates. By using LiDAR-based measurements of crown width (dcrm), the RMSE was 46%.

22. RESULTS – Forest inventory Omission errors → 10% of total volume missed DBH underestimation by 1 cm → underest. of total and saw wood volume DBH-estimates were averaged → skew in saw/pulp wood proportions

23. Conclusions The system worked partly well, and it solved all of the STRS-tasks, and provided results per species and timber sortiment, but - Results of timber resources were contaminated by large systematic errors and noise due to measurement errors and model errors (averaging)  Calibration of measurements and model estimates is needed, also a better allometric DBH-estimation phase - Many of the tasks need a more automatic solution, especially the Sp-recognition task.