cartus_th4.t02.3.ppt
TRANSCRIPT
Ecosystem Structure Measurements from DESDynI:Technological options and data fusion using Small-Footprint Lidar
and ALOS/PALSAR data over Central Chile
Josef KellndorferScott GoetzWayne WalkerOliver Cartus
The Woods Hole Research Center(www.whrc.org)
Markus Rombach, Digimapas Chile
Sergio Gonzalez, Arauco Timber Company
Ralph Dubayah, University of Maryland
(1) For showcase scenario of Central Chile: Evaluate multi-sensor data-fusion strategies for canopy height (CH) and Growing Stock Volume (GSV) retrieval.
(2) Simulate DESDynI Radar (initially also Lidar) performance: Development of DESDynI Radar ecosystem structure retrieval algorithms
(3) Increasing availability of small-footprint Lidar for forest structure analysis development of algorithms for up scaling in situ measurements to airborne samples (Lidar) and space-borne (SAR, optical data) wall-to-wall data
(4) Investigation of SAR saturation, forest structural differences, benfit of having multi-temporal SAR data, InSAR coherence
Study ObjectivesStudy Objectives
Chile- ALOS FBD Data MosaicChile- ALOS FBD Data Mosaic
ALOS Processing 189 SLC Products (FBD) ordered from
ASF Data for 2007,2009,2010 Multi-temporal coverage: 1-3 images
per year GAMMA Speckle Filter Terrain-corrected geocoding with
SRTM-3 DEM Resampled to 15x15 m grid Topographic Normalization
First study area:
Highly managed plantations of: radiata pine (Pinus radiata)bluegum eucalyptus (Eucalyptus globulus)
Landsat ETM+ dataLandsat ETM+ data
Global Land Survey 2005Global Land Survey 2005Mosaic 1. Dec. 2005/ 30. Jan. 2007Mosaic 1. Dec. 2005/ 30. Jan. 2007
(Dry season)(Dry season)
Downloaded from USGS Earth Explorer: http://earthexplorer.usgs.gov. Product Type: L1T, i.e. data is terrain corrected.SLC gaps filled for ~98%Metrics: Reflectances, Tasseled Cap Transformations and NDVI
Laser scanning system: Riegl LMS-Q560. Range capture: Full waveform digitization. Field of view: ± 22,5 degrees Laser wavelength: 1550nm Operating altitude: 30 m - 800 m AGL Beam divergence: 0.5 mrad Swath width: 80 % of op. altitude (45 degrees) Range resolution: 0.020 m Vertical accuracy: < 0.15 m Horizontal accuracy: < 0.25 m
DIGIMAPAS: Airborne Small Footprint Lidar data
1m Canopy Height Model, CHM1m Canopy Height Model, CHM
CHM (1 m pixels) = DSM (first return) – DTM (last return)
DTM DSM CHM
ARAUCO Stand Level Inventory data for 2007
CHMInventory data
Database provides Canopy Height (100 highest trees), DBH, Growing Stock Volume (GSV m3/ha), Species, etc. for ~7000 standsAccording to ARAUCO data is representative for 100 000 ha of plantation forest For now: 440 stands (mainly radiata pine stands)
Extraction of Lidar Metrics for inventory standsExtraction of Lidar Metrics for inventory stands
Histogram-simulated Lidar waveforms from 1-meter small-footprint lidar data:
Examples show simulated waveforms (resp. canopy height profiles) at hectare scale stand level.
Extracted metrics:-Relative Heights 10-100 (percentiles of CHM distribution)-Canopy Cover (% pixels > 2 or 5 m)-Number of Gaussians
Lidar Metrics vs. in situ dataLidar Metrics vs. in situ data
In particular RH30 - RH90 well correlated to GSV and canopy height (Height 100)RH100 differs noticeably from in situ tree heightRelative Heights highly correlated (do they provide different information)
Lidar RetrievalLidar Retrieval
Modeling with RandomForest ensemble regression trees:
-Straightforward approach for combining various types of data
-Provides tools for analyzing the importance of certain predictors (e.g. RH)
-Accuracy assessment: all presented numbers will be so-called out-of-bag estimates (bootstrap validation)
Lidar Retrieval (stands > 2ha)Lidar Retrieval (stands > 2ha)
Topography did not affect retrieval performance. Retrieval did not improve when integrating canopy cover or number of Gaussians
Retrieval based on empirical relationship between: RH90 (best predictor for height), Height and GSV (25 % Training and 75 % test samples)
Retrieval performance only slightly lower than in case of retrieval with RandomForest and various Lidar metrics
The high correlation between Height and GSV appears to be the main driver for the Lidar-based GSV retrievalGSV/Height Allometry depends on …
GSV/Height Allometry
Relative StockingRelative Stocking
Relative stocking, RS: ratio of observed basal area at a certain stand age and an optimal basal areaDifferences in RS because of different site or seedling quality
GSV retrieval accuracy (with RandomForest) higher for high relative stocking stands, e.g.:
RMSEr GSV = 18% (RS>75%)RMSEr GSV = 23 % (RS<60%)
RS overall high in test area (natural forests?)
Effect of Footprint Size Effect of Footprint Size
Footprints of 100 - 20 000 m2 extracted from CHM within in situ stands
DE
SDY
NI footprint
Even with 10 m footprints Lidar metrics are representative for the entire hectare scale stand
Synergy:Synergy:Lidar, ALOS PALSAR, Landsat ETMLidar, ALOS PALSAR, Landsat ETM
Approach:Lidar GSV/Height estimates are used as
response variables to train RandomForest Models with ALOS/ETM data as
predictors
Test for ~10 km large subset
3 x HH/HV intensity Landsat ETM+
1m Lidar Canopy Height Model
1)eCognition segmentation2)GSV/Height retrieval
High Multi-temporal consistency of HH & HV backscatter
ALOS Images:05 July 2007 (dry)05 Oct. 2007 (rainy)21 Oct. 2007 (dry)
18
When using only 1 ALOS FBD HH/HV image pair (stands > 2 ha)
- Predictors in RandomForest: HH & HV Intensity + Coefficient of Variation CV (texture)
- Differences in Retrieval Accuracy between the three images only 2 % although they were acquired under different weather conditions
3 ALOS HH/HV images (stands > 2 ha)
Two multi-temporal approaches were tested: 1)All images as predictors in one model (upper row)2)Separate model for each FBD image weighted multi-temporal combination (cf. Santoro et al., 2006) (lower left plot)
Integration of multi-temporal data improves retrieval performance
3 ALOS HH/HV images + ETM+ (stands > 2 ha)
Multi-temporal combination of single image estimates
3 ALOS HH/HV images + ETM+ (stands > 5 ha)
GSV<300 m3/ha
Summary:
1m Lidar CHMs allowed GSV/Height retrieval with relative errors of 23 and 7 % respectively (RMSE of 63 m3/ha and 1.7 m) for highly managed pine plantations. For high relative stocking stands accuracy even higher.
Use of Random Forest with various Canopy metrics hardly improves the retrieval performance compared to allometric relations based on single metrics
When using Lidar GSV/Height estimates to train models for ALOS/ETM data:
Availability of multi-temporal data improved the retrieval (5-10%)
Integration of ETM data improved the retrieval
Outlook: Extrapolation of ALOS/ETM+ models/retrieval to large areas / entire Chile
How many Lidar transects are required to capture the spatial variability of backscatter signatures over forest ?
Retrieval performance was assessed for highly managed plantation forests. Accuracy in other natural forests?
How much can repeat-pass coherence contribute (reliably?) to large area GSV/Height retrieval - repeat-pass coherence strongly depends on weather conditions and the baseline
Summary and OutlookSummary and Outlook
Stands >2 ha
ALOS HH& HV intensity vs. Lidar GSV estimates
Images:
05 July 2007 (dry)05 Oct. 2007 (rainy)21 Oct. 2007 (dry)
24
ETM+ only (stands > 2 ha)
Predictors:ETM bands 1-5Tasseled Cap transformationsNDVI Accuracy similar to that
achieved when using 1 ALOS HH/HV image pair