USING LIDAR AND RAPIDEYE TO PROVIDE

Transcription

1 USING LIDAR AND RAPIDEYE TO PROVIDE ENHANCED AREA AND YIELD DESCRIPTIONS FOR NEW ZEALAND SMALL-SCALE PLANTATIONS Cong (Vega) Xu Dr. Bruce Manley Dr. Justin Morgenroth School of Forestry, University of Canterbury

2 BACKGROUND AND INTRODUCTION Small-scale plantation forests (30% of all plantations) are not well understood in net stocked area Small-scale forests lacks yield information NZ lacks accurate spatial representation of small-scale plantations Increasing availability of cost-effective remote sensing data

3 RESEARCH OBJECTIVES Evaluate different combinations of remote sensing techniques and datasets in mapping net stocked plantation forests Evaluate different modelling approaches and remote sensing datasets in modelling height, basal area, volume and stand age Apply the selected area mapping and modelling approaches to the Wairarapa region

4 REMOTE SENSING DATASETS Dataset Resolution Temporal Coverage Description Application Aerial Photography 0.3 m Dec Jan 2013 Orthorectified aerial photography: RGB Ground truthing for forest mapping Airborne LiDAR 3.7 points m -2 Jan-Dec 2013 Wall-to-wall for Wellington Region Derived surfaces for forest mapping, metrics for model stand variables RapidEye 5 m Nov Feb band multispectral imagery: RGB, RE, NIR Derived surfaces for forest mapping, metrics for model stand variables

5 AREA SAMPLE SELECTION FOR MAPPING

6 AREA PROCESS OVERVIEW NN: Nearest Neighbour CART: Classification and Regression Tree

7 AREA RESULTS CLASSIFICATION ACCURACY OF DIFFERENT MAPPING APPROACHES AND DATASETS 100% 90% 80% 80% 82% 81% 88% 75% Classification Accuracy 70% 60% 50% 40% 30% 20% 10% 60% 63% 67% 0% NN- RE only NN- RE+LiDAR CART-RE only CART- RE+LiDAR Plantation Overall

8 AREA - RESULTS FOR PLANTATION ALL VALIDATION GRIDS 79% 91% 89% 91% All Producer's accuracy Excl. temporal difference User's accuracy Total Digitised Total Mapped Difference Difference MAE RMSE (ha) (ha) (ha) % (ha) (ha) All standing trees % Exclude new plantings % Note: New plantings are generally not visible on satellite imagery

9 AREA VISUAL COMPARISON

10 AREA - PATCH-LEVEL COMPARISON 423 sets of valid patch to patch comparisons All patched: Average patch size: 9.5 ha, mean absolute error = 0.8 ha Large areas are more accurately mapped

11 MODELLING STAND VARIABLES -PLOT SUMMARY Forest Measurement Approach (FMA) plots Pre-harvest inventory 112 plots Stand Variables Mean Range Plot Area (ha) Slope ( ) Age (years) Stocking (stems ha -2 ) Diameter at Breast Height (mm) Individual Tree Height (m) Mean Top Height (m) Basal Area (m 2 ha -1 ) Volume (m 3 ha -1 )

12 MODELLING STAND VARIABLES -APPROACH Input predictors: LiDAR (111): height, canopy and intensity metrics RapidEye (68): spectral, textural and vegetation indices Parametric models Multiple Linear Regression (MLR) Seemingly Unrelated Regression (SUR) Non-parametric models K Nearest Neighbour (knn) Random Forests (RF) 10-fold cross-validation RMSE = Σ 2 n MD= Σ( ) n

13 MODELLING STAND VARIABLES MODEL COMPARISON BASED ON 10-FOLD CROSS-VALIDATION Comparison of Root Mean Square Error as a percentage of predicted mean (RMSE%) for MTH, BA, VOL and age estimated by MLR, SUR, k-nn and RF models.

14 MODELLING STAND VARIABLES - BEST MODEL Best model for each stand variable: (lowest RMSE) Stand Variable Model Input Data RMSE (RMSE%) MD (MD%) MTH (m) BA (m 2 ha -1 ) VOL (m 3 ha -1 ) Age (years) RF LiDAR 1.37 (5.4%) 0.05 (0.19%) MLR LiDAR + RapidEye 9.42 (18.54%) 0.24 (0.47%) MLR LiDAR + RapidEye (19.71%) 1.67 (0.36%) knn LiDAR + RapidEye 2.05 (10.53%) 0.02 (0.12%) Best single model- MLR with LiDAR metrics Stand Variable Input Data RMSE (RMSE%) MD (MD%) MTH (m) BA (m 2 ha -1 ) VOL (m 3 ha -1 ) Age (years) LiDAR 1.81 (6.9%) 0.01 (0.04%) LiDAR 9.92 (19.54%) 0.24 (0.47%) LiDAR (20.46%) 2.95 (0.64%) LiDAR 2.17 (11.17%) 0.07 (0.35%)

15 APPLICATION TO WAIRARAPA -APPROACH Area 21 RapidEye scenes and LiDAR surfaces Automated CART classification Manual mapping of young plantation Stand variables Derive 5 x 5m LiDAR metrics Estimate MTH, BA, VOL and age using MLR Calculate mean for each polygon

16 PLANTATION AREA

17 MTH

18 BA

19 VOL

20 AGE

21 APPLICATION TO WAIRARAPA Plantation Area Mapped Digitised young Total plantation NEFD LCDB plantation (ha) plantation (ha) (ha) plantation (ha) plantation (ha) Total Stand variables Stand Variable Input Data RMSE (RMSE%) MD (MD%) Stand Variable MTH (m) BA (m 2 ha -1 ) VOL (m 3 ha -1 ) Age (years)

22 APPLICATION TO WAIRARAPA AGE-CLASS DISTRIBUTION

23 APPLICATION TO WAIRARAPA -YIELD COMPARISON

24 CONCLUSION Best mapping approach: Combined RapidEye and LiDAR with CART Best modelling approach: MLR using LiDAR metrics Wairarapa application Fails to detect young plantings (6%) 3.4% lower than NEFD 287 ha higher than UC 2017 Case study (0.6%) 25 m 3 ha -1 lower than WAF yield

25 IMPLICATION Improve understanding of small-scale forests Identify where they are What the productive areas are How much wood is there Application to all regions in NZ Develop a national geospatial database of plantation Estimate stand variables for the plantations Allow future update and monitoring of the resources

26 ACKNOWLEDGEMENT School of Forestry PhD Scholarship Blackbridge RapidEye Landcare Research Land Information New Zealand Wellington Regional Council Indufor Asia Pacific Michael Watt (Scion) Jonathan Dash (Scion) Huimin Lin (School of Forestry) Dr Luis Apiolazas and Dr Daniel Gerhard (UC) Alan Bell and forest managers in Wairarapa

27 THANK YOU