Multisensoral UAV-Based Reference Measurements for Forestry Applications Research Manager D.Sc. Anttoni Jaakkola Centre of Excellence in Laser Scanning Research
2 Outline UAV applications Reference level UAS measurements Low-cost LiDAR High-end LiDAR Hyperspectral camera FMCW radar Multrispectral LiDAR Future outlook of UAV LiDAR
History 3
4 Strengths of UAS Measurements Small area mapping Corridor mapping Multitemporal measurements Air/spaceborne sensor Simulation Validation Reference data collection
Alternative Platforms 5
Mapping: Urban Planning 6
Corridors: Roads 7
Corridors: Power Lines 8
Corridors: Rivers 9
Multitemporal Data 10
UAV-Based Reference Data Collection 11 Evo test site Southern Finland Boreal forest Pine Spruce Birch 91 test plots 32x32 m
Low-Cost LiDAR 12 Velodyne VLP-16 Lite Novatel SPAN-IGM S1 40 m AGL 10 m line spacing 8 m/s Up to 800 pts/m 2 Single plot
Low-Cost LiDAR 13 Velodyne Puck LITE Novatel IGM-S1 Size Ø103 x 72 mm Size 152 x 142 x 51 mm Weight 590 g Weight 540 g Measurement range 100 m Measurement rate 125 Hz Pulse repetition rate 300 000 Hz Horizontal accuracy 1 cm Profile frequency 16 x 20 Hz Vertical accuracy 2 cm Range accuracy ±3 cm Roll/Pitch accuracy 0.015 Beam divergence 3 mrad Heading accuracy 0.080
Low-Cost LiDAR 14
Low-Cost LiDAR 15 Bias Bias (%) RMSE RMSE (%) R Tree height (m) 0.02 0.08 1.02 5.16 0.92 DBH (cm) 0.02 0.07 2.55 10.40 0.88 Basal area (m 2 ) 0.00 0.56 0.01 19.73 0.84 Volume (m 3 ) 0.00-0.02 0.09 19.26 0.88 Biomass (Mg) 0.12 0.05 40.81 17.35 0.89
16 High-End LiDAR Riegl VQ-480-U Novatel SPAN-LCI 75 m AGL Single overpass >100 pts/m 2
17 High-End LiDAR Riegl VQ-480-U Size Weight Measurement range Pulse repetition rate Profile frequency Range accuracy Beam divergence Ø183 x 348 mm 7.5 kg 200-1500 m 550 000 Hz 150 Hz ±2.5 cm 0.3 mrad
High-End LiDAR 18
High-End LiDAR 19
High-End LiDAR 20
21 High-End LiDAR Non-calibrated Calibrated Height DBH Basal area Volume Biomass Bias (%) 0.28-4.56 1.06 0.19 0.57 RMSE (%) 4.54 8.88 16.91 17.14 16.77 R 0.97 0.97 0.81 0.90 0.88
22 Hyperspectral Camera Fabry-Pérot interferometric hyperspectral camera RGB camera GNSS L1 receiver 90 m AGL 4 m/s
23 Hyperspectral Camera Rikola/Senop FPI camera 500-900 nm 1010 x 1010 pixels Up to 30 fps Samsung NX300 APS-C, 24 mpix
Hyperspectral Mosaics 1 2 3, 4 5 8 6 7 9 10 11
Individual tree classification Reference data: - fieldwork data (location & species) Classification features 3D features of individual trees (RGB pointcloud) Lastools Classification Feature Selection Classification training and evaluation Weka Spectral features from hyperspectral FPI mosaics Matlab Classification of detected trees Photoscan Photogrammetric point cloud from RGB imagery FUSION Software (Pacific Northwest Research Station) ) Digital Surface Model (DSM) Digital terrain model (DTM) from NLS ALS data Canopy Height Model (CHM) = DSM DTM Individual tree detection using local maxima method
Hyperspectral Camera 26 True Class Precision Pine Spruce Birch Larch Classified as Recall Pine Spruce Birch Larch 2584 41 0 2 0.984 122 692 2 6 0.842 13 8 558 1 0.962 11 1 5 105 0.861 0.947 0.933 0.988 0.921 All Features Spectral Features Feature Selection All Features (no norm. spectra) Spectral Features (no norm. spectra) Feature Selection (no norm. spectra) 3D Features Accuracy (%) 95.11 94.65 94.89 94.84 94.29 92.58 72.01 Kappa 0.91 0.90 0.90 0.90 0.89 0.86 0.39
27 FMCW Radar Frequency 14.0 GHz (Ku) Sweep frequency 1000 Mhz Range resolution 15 cm Beam width 6 Polarization HH, VV, HV, VH Weight 14 kg (UAV version 6 kg) Single line, 60-80 m AGL
FMCW Radar 28
FMCW Radar 29
30 FMCW Radar (a) B E db db A (b) C D
FMCW Radar 31
32 FMCW Radar 35 30 25 mean height (m) 2096 50 40 mean DBH (cm) Observed 20 Observed 30 15 20 Mean height (m) Mean DBH (cm) Basal area (m 2 /ha) Volume (m 3 /ha) Biomass (Mg/ha) Bias Bias (%) RMSE RMSE (%) R -0.09-0.42 2.83 12.99 0.78-0.02-0.07 5.59 21.25 0.65 0.10 0.35 5.89 21.34 0.61-0.80-0.28 67.19 23.61 0.77-0.19-0.14 33.05 23.39 0.69 Observed Observed 1041 10 10 15 20 25 30 10 20 30 40 Predicted Predicted 50 2 basal area (m /ha) 40 30 20 10 0 10 20 30 40 Predicted 250 biomass (Mg/ha) 200 150 100 Observed 600 3 /ha) volume (m 500 400 300 200 100 0 100 200 300 400 500 Predicted 50 0 50 100 150 200 Predicted
33 Multispectral LiDAR Optech Titan
34 Multispectral LiDAR 155 150 145 140 135 Chan1 Chan2 Chan3 155 150 145 140 135
35 Multispectral LiDAR Confusion matrix based on intensity features Reference Predicted Producer Pine Spruce Birch Pine 623 12 16 95.70 Spruce 32 180 27 75.31 Birch 47 18 197 75.19 User 88.75 85.71 82.08 Overall = 86.81%
36 Multispectral LiDAR Confusion matrix based on point cloud and intensity features Reference Predicted Producer Pine Spruce Birch Pine 622 14 15 95,55 Spruce 18 201 20 84,10 Birch 46 21 195 74,43 User 90.67 85.17 84.78 Overall = 88.36%
37 Future Outlook of LiDAR Technology Automotive industry Solid state / flash LiDAR technology Multispectral LiDAR Single photon LiDAR
38 Automotive Industry Driving the development of low-cost light-weight LiDARs Range accuracy or beam divergence are not the primary targets
39 Solid-State Technology Automotive applications require low cost and high reliability Optical beam steering
40 Flash LiDAR Low volume and high price of InGaAs components Frame based, but low resolution
41 Multispectral LiDAR Optech Titan or multiple scanners, e.g. Faro S120 + X330 Currently heavy and expensive
42 Single photon LiDAR Currently available for full-scale airborne laser scanning
43 Summary UAV-based measurements can be used as reference data for some applications Efficiency compared to manual measurements may be higher by an order of magnitude Current LiDAR sensors are still heavy, expensive, inaccurate and/or limited in measurement range Development towards new sensor technologies is rapid