LiDAR Product Suite for the Hearst Forest

Transcription

1 LiDAR Product Suite for the Hearst Forest & other R&D activities present and future Murray Woods Ontario Ministry of Natural Resources Doug Pitt Canadian Wood Fibre Centre

2 Advanced Forest Resource Inventory Technologies

3 TEAM AFRIT Doug Pitt Dave Nesbitt Margaret Penner Forest Analysis Ltd. Kevin Lim Lim Geomatics Paul Treitz Dave Etheridge Jeff Dech Don Leckie François Gougeon

4 National CWFC Program Regional Focus: AFRIT-Hearst

5 Outline LiDAR 101 LiDAR Derived Inventory Current LiDAR Inventory Products Other Hearst R&D efforts Looking forward

6 LiDAR Light Detection And Ranging Discrete System Active remote sensing technology; transmit & receive ~35, ,000 pulses of NIR laser light per second GPS provides the exact X-Y-Z position of each return Discrete System - Each pulse can produce multiple returns (3-5) Full Waveform System not limited to number of returns but detectible signals Used with permission of Doug Pitt

7 LiDAR Light Detection And Ranging Discrete System Active remote sensing technology; transmit & receive ~35, ,000 pulses of NIR laser light per second GPS provides the exact X-Y-Z position of each return Discrete System - Each pulse can produce multiple returns (3-5) Full Waveform System not limited to number of returns but detectible signals 2007 LiDAR

8 LiDAR Light Detection And Ranging Full Waveform System Active remote sensing technology; transmit & receive ~35, ,000 pulses of NIR laser light per second GPS provides the exact X-Y-Z position of each return Discrete System - Each pulse can produce multiple returns (3-5) Full Waveform System not limited to number of returns but detectible signals Source: e-education.psu.edu

9 LiDAR Light Detection And Ranging Full Waveform System (~48 rts/m 2 ) Active remote sensing technology; transmit & receive ~35, ,000 pulses of NIR laser light per second GPS provides the exact X-Y-Z position of each return Discrete System - Each pulse can produce multiple returns (3-5) Full Waveform System not limited to number of returns but detectible signals 2012 LiDAR

10 LiDAR Light Detection And Ranging Discrete System (~3 rts/m 2 ) Full Waveform System (~48 rts/m 2 ) Active remote sensing technology; transmit & receive ~35, ,000 pulses of NIR laser light per second GPS provides the exact X-Y-Z position of each return Discrete System - Each pulse can produce multiple returns (3-5) Full Waveform System not limited to number of returns but detectible signals 2007 LiDAR 2012 LiDAR

11 LiDAR Light Detection And Ranging DSM DEM Imagery 11 3

12 LiDAR Light Detection And Ranging DSM DEM CHM Imagery 12 3

13 LiDAR Light Detection And Ranging Historical Reason for Acquiring LiDAR Data OBM 20m LiDAR 1m Classified Ground Returns Prince Edward County, Ontario

14 LiDAR Technology Classified non-ground Returns Classified Ground Returns Prince Edward County, Ontario

15 Operational Cruising Field Sampling Cruising generally not occurring in Boreal forests Very light 1%-2% Expensive only provides information on sampled sites extrapolated Need to do for the next stand and the next

16 LiDAR Derived Inventory Field Sampling vs. 100% Enumeration 100% enumeration of the landbase with LiDAR measurements of vertical structure Scalability permits the use of regression estimators to scale PU estimates to groups of PUs, Stand, Block or Forest

17 LiDAR Derived Forest Inventory - Approach 2010 Field Program Established 446 Calibration 400m 2 plots 64 Validation plots (2012) Sampled 8 Forest Types x Development stages LC SB SF SP PJ IH MWC - MWH All trees measured for DBH and height sample > 10 cm Trees <10 cm density count by 1 m height classes avg DBH Ecosite description

18 LiDAR Derived Forest Inventory - Approach Additional forest inventory information contained in point cloud data Focus of AFRIT is Area based modeling NOT Individual Tree Prediction Unit = 20m X 20m (400m 2 ) vegetation ground

19 LiDAR Derived Forest Inventory - Approach Additional forest inventory information contained in point cloud data Focus of AFRIT is Area based modeling NOT Individual Tree Prediction Unit = 20m X 20m (400m 2 )

20 LiDAR Derived Inventory Pairing with Ground Data Field Plot Measurement

21 Enhanced Modeling Approaches - Simplified Regression randomforest y = b o + b 1 x 1 + b 2 x 2 Option for hundreds to thousands of trees

22 Comparison of Basal Area Predictions from RF and SUR Hearst Forest Observed RF SUR_FU 50 Basal Area (m 2 /ha) LC MWC MWH PJ IH SB SF SP All Forest Type

23 Enhanced Modeling Approaches Regression RandomForest Requires an existing polygon inventory Requires separate forest-type models to statistically be built - specialist Can t adjust to within polygon species variation Can extrapolate beyond model building data doesn't rely on an existing polygon inventory no separate forest-type models to statistically build very quick to implement custom application of LiDAR predictions for each 20m x 20m Predictions are equivalent to Regression

24 LiDAR Derived Inventory LiDAR predictive Models for: Height (AVG, Top) DBHq (~avgdbh) Volume (GTV, GMV) Basal area Biomass Density* Sawlog Volume Close Utilization Volume Dom/Codom Ht Mean Tree GMV Size Class Distributions Predicted LiDAR - SUR * Derived from DBHq & BA Observed All Data

25

26

27

28 Dominant/Codominant Ht

29 DBHq

30 Gross Merchantable Volume

31 Gross Merchantable Volume m 3 /ha

32 Block Summary 186 ha a) 97.5 m 3 /ha * 186 ha = 18,135.0 m 3 75 ha b) m 3 /ha * 75 ha = 11,587.5 m m 3 /ha * 261 ha = 29,701.8 m 3 Minimum: 0.75 m 3 /ha Maximum: m 3 /ha

33 Optimistic Sawlog Volume Mean Tree Volume Derived Pulp Volume Merchantable Basal Area

34 Hearst LiDAR Derived Size-Class Distributions Black Spruce Validation Data by VCI class Size Class Distributions from Airborne LiDAR for a Boreal Forest. Woods et al 2013 In Prep. Van Ewijk, K.Y, P.M. Treitz and N. A. Scott Characterizing forest succession in central Ontario using Lidar-derived indices. Photogrammetric Engineering and Remote Sensing 77:

35 i

36

37

38

39

40 Gross Merchantable Volume

41 LiDAR Derived Inventory RMF Block Validation Pj Clearcuts: ~within 2-5% Po Clearcuts: ~ within 5 15% Sb with retention: ~5-25% net down for cull Retention has to be factored in Image and data courtesy Tembec Inc.

42 Approach Volume Comparison LiDAR provides better volume prediction (statistically tested) Yield Curve Yield Curve: 14.5% LIDAR: 5.4 % BW SP PJ PO % 10% 0% -22% % 33% 3% -34% % -15% 0% 1% 244-9% 34% -12% -14% 245-2% 9% -1% -6% 246-2% 20% -1% -17% 251 6% 33% 6% -45% 252 9% 55% 1% -66% 254-1% 4% 10% -12% 256 1% -6% 5% 257 7% -17% 9% 259 1% 0% -1% 275 Difference from harvested volume (scaled) LIDAR BW SP PJ PO 208 2% -6% 0% 3% % 6% 1% 8% 220 0% -11% 4% 7% % 2% 2% 5% 245 0% 5% 1% -6% 246 0% 5% 1% -6% 251 1% 11% 5% -17% 252 4% 23% 2% -29% 254 1% -4% 0% 4% 256 2% -1% -1% 257 2% -5% 3% 259-3% -2% 5% % & under greater than 10 %

43 Hearst Forest ITC Individual Tree Classification

44 Hearst Forest ITC Individual Tree Classification

45 Hearst Forest ITC Individual Tree Classification ~150,000 ha operationally classified with ITC 730 stands Range of Forest Types Range of Height Classes 2 Certified Forest Interpreters vs. Computer ITC Software Misclassification type Leading Species Comparison Interpreter 1 vs. 2 Interpreter vs. ITC Agreement 59% 39% Close (similar species e.g., Sw, Sb) 71% 51% Within hardwood or softwood 91% 81% Serious disagreement 9% 19%

46 Hearst Forest ITC Individual Tree Classification Interpreter 1 Interpreter 2 % Ab Bf Bw Cw La Pb Pj Pt Sb Sw agreement Ab 0% * Not significantly different Bf % Bw % Cw % La % Pb % Pj 0% Pt % Sb % Sw % Total % Interpreter ITC % Ab Bf Bw Cw La Pb Pj Pt Sb Sw agreement Interp 1 vs. Interp 2: % Agreement for NE SFU - 55% Ab 2 1 0% Bf % Bw % Cw % La % Pb % Pj 1 1 0% Pt % Sb % Sw % Total %

47 Photo interpreter height (m) Photo Interpreter 2 height (m) Hearst Forest ITC Individual Tree Classification Height a) * Not significantly different OHT_2 10 1:1 line Photo Interpreter 1 height (m) b) * Significantly different for a reason OHT_2 1:1 line OHT_ ITC height (m)

48 Photo interpreter height (m) Photo Interpreter 2 height (m) Hearst Forest ITC Individual Tree Classification Crown Closure a) * Significantly different OHT_2 10 1:1 line b) Photo Interpreter 1 height (m) * Significantly different OHT_2 1:1 line OHT_ ITC height (m)

49 Ongoing Enhancements Low Density LiDAR based Species Group Classification Predicted (%) Actual HWD MWC MWH CONIFER HWD MWC 100 MWH CONIFER 4 96 N = 346 plots; 86 used for validation Woods et al. (in prep)

50 Ongoing Enhancements LiDAR Species Classification from Low to High Density Error matrix of classification Baoxin Hu Dept. of Earth and Space Science and Engineering York University, Toronto, Canada

51 Ongoing Enhancements (Akumu et al. In Prep) Environmental input variables: Elevation (10m) Slope (%) Surface shape (Curvature) Mode of deposition (NOEGTS) Landcover Slope Position from TPI (macro window = 1km) (medium window = 500m) (micro window = 20m) Wetness Index

52 Ongoing Enhancements Wet Areas Mapping Site Productivity Linkage - Ecosite

53 Ongoing Enhancements WAM Mapping Site Productivity - Ecosite

54 Ongoing Enhancements WAM Mapping Site Productivity - Ecosite

55 Ongoing Enhancements Fibre Analysis Jeff Dech, Bharat Pokharel, Art Groot Sampling on AFRIT- Hearst Plot Network

56 Looking Forward Semi Global Matching - Pixel Correlation Not the same as LiDAR but similar Image based product Very detailed DSM Lower cost acquisition produced with each inventory period LiDAR SGM

57 Study on the Hearst Forest to: Average stem characteristics 100 Develop process to utilize SGM point cloud like LiDAR Observed LiDAR-predicted SGM-predicted with LiDAR DEM m 3 /100 Compare potential SGM predictions with those derived from LiDAR m m cm 0 Used the LiDAR derived Hearst DEM to normalize SGM data Top Height D-C Height QM-DBH AGMV Stand characteristics m 2 /ha m 3 /ha x 10 m 3 /ha x stems/ha x A comparison of airborne LiDAR and digital photogrammetric point clouds for the areabased estimation of forest inventory attributes in boreal Ontario Doug G. Pitt, Murray Woods, and Margaret Penner (Submitted for review November 2013) BA GMV SLV Density

58 Top height (m) LiDAR SGM-LiDAR_DEM Stem Characteristics D-C height (m) QM-DBH (cm) RMSE = 1.21 m RMSE = 1.69 m RMSE = 2.11 cm RMSE = 1.54 m RMSE = 1.92 m RMSE = 2.57 cm Forest Type SB (34%) SP (25%) SF (5%) PJ (3%) LC (3%) MWC (18%) MWH (7%) PO (5%) image-based point clouds were found produce accuracies that were equivalent to LiDAR, however some losses in precision were evident (Pitt et al in review) AGMV (m 3 ) RMSE = 0.05 m RMSE = 0.06 m

59 BA (m 2 /ha) LiDAR SGM-LiDAR_DEM Stand Characteristics GMV (m 3 /ha) SLV (m 3 /ha) RMSE = 7.47 m 2 /ha RMSE = 38.7 m 3 /ha RMSE = 8.00 m 2 /ha RMSE = 38.4 m 3 /ha + Forest Type SB (34%) SP (25%) SF (5%) PJ (3%) LC (3%) For variables such as basal area, gross merchantable volume and sawlog volume, such losses in precision amounted to less than 10% increases in mean absolute errors RMSE = 35.0 m 3 /ha RMSE = 36.0 m 3 /ha MWC (18%) MWH (7%) PO (5%) (Pitt et al in review) Density (stems/ha) RMSE = 421 stems/ha RMSE = 468 stems/ha

60 SGM Predictions Normalized on OBM 20m DEM Increased loss of precision from process using LiDAR DEM May still be an enhancement CAUTION Study area was relatively flat Top height (m) D-C height (m) QM-DBH (cm) AGMV (m 3 ) RMSE = 1.82 m RMSE = 1.82 m RMSE = 2.98 cm BA (m 2 /ha) GMV (m 3 /ha) SLV (m 3 /ha) Density stems/ha RMSE = 7.99 m 2 /ha RMSE = 46.3 m 3 /ha RMSE = 43.5 m 3 /ha Forest Type SB (34%) SP (25%) SF (5%) PJ (3%) LC (3%) MWC (18%) MWH (7%) PO (5%) (Pitt et al in review) 0.1 RMSE = 0.07 m RMSE = 506 stems/ha

61 FPInnovations Advantage Report 14(1), 2013, Pointe-Claire, QC, 8 p.

62 Objectives Are LiDAR-enhanced inventories leading to better forest management decisions? What are the cost savings associated with enhanced decision making?

63 Approach Volume Comparison LiDAR provides better volume prediction (statistically tested) Yield Curve Yield Curve: 14.5% LIDAR: 5.4 % BW SP PJ PO % 10% 0% -22% % 33% 3% -34% % -15% 0% 1% 244-9% 34% -12% -14% 245-2% 9% -1% -6% 246-2% 20% -1% -17% 251 6% 33% 6% -45% 252 9% 55% 1% -66% 254-1% 4% 10% -12% 256 1% -6% 5% 257 7% -17% 9% 259 1% 0% -1% 275 Difference from harvested volume (scaled) LIDAR BW SP PJ PO 208 2% -6% 0% 3% % 6% 1% 8% 220 0% -11% 4% 7% % 2% 2% 5% 245 0% 5% 1% -6% 246 0% 5% 1% -6% 251 1% 11% 5% -17% 252 4% 23% 2% -29% 254 1% -4% 0% 4% 256 2% -1% -1% 257 2% -5% 3% 259-3% -2% 5% % & under greater than 10 %

64 Approach Cost Analysis Comparison of 2 scenarios 1. Actual cut over ACTUAL Area harvested within approved plan 2. Redesign blocks within proposed plan LIDAR New harvest planning using LiDAR for area similar to initial plan Validated by Tembec FRM Stay within approved 5 year harvest blocks

65 Approach Cost Analysis: Scenarios Redesign plan, based on full suite of LiDAR data products Removed area of less GMV from Scenario DBHq (cm) ACTUAL GMV (m 3 ha) LIDAR

66 Approach Cost Analysis Categories Three groups of cost items : A. Inventory acquisition and processing - 6 items B. Forest operations 20 items C. Mill - 4 items

67 Results: Actual vs. LiDAR Scenario Three groups of cost items : A. Inventory acquisition and processing - 6 items B. Forest operations 20 items C. Mill - 4 items $ m 3 $ 1.40 m 3 $ 0.30 m 3 Savings $ 1.60 m 3 X 500,000 m 3 / year = $ 800,000 / year Payback = 1.3 years *Tembec s LiDAR acquisition 50% of what was modeled so payback was achieved in the first year!!!

68 July/Aug Vol 86. No 4 Forestry Chronicle Remote Sens. 2013, 5, ; doi /rs FPInnovations Advantage Report 14(1), 2013, Pointe- Claire, QC, 8 p. Contact information: murray.woods@ontario.ca dpitt@nrcan.gc.ca