benson_we3t051.pdf

MotivationIntroduction

ModelsApplication to BOREAS Data

Classification AlgorithmResults and Conclusions

Forest Structure Estimation in the CanadianBoreal forest

Michael L. Benson Leland E.Pierce Kathleen M. BergenKamal Sarabandi Kailai Zhang Caitlin E. Ryan

The University of Michigan, Radiation Lab & School of Natural Resources andthe Environment

Ann Arbor, MI 48109-2122 USA

Benson, et al. IGARSS 2011 Forest Structure Estimation using SAR, LiDAR, and Optical data in the Canadian Boreal forest




Goal: Accurate estimation of Forest Structure parametersusing measured SAR, LIDAR, and Optical data.

Motivation: Forest Structure is important ecologically for globalclimate estimation as well as biodiversity and othertopics.

This Talk: Use a set of simulators for each sensing modality aswell as real remotely sensed data and presents aninversion algorithm capable of accurate forestparameter retrieval requiring a minimal amount ofancillary / ground truth data.





Outline

1. Introduction

2. Background

3. Approach

4. Forward Models & Database GenerationI Forest Geometrical ModelI Optical ModelI SAR ModelI LIDAR Model

5. Application to BOREAS Remotely Sensed Data sets

6. Classification Algorithm

7. Results

8. Conclusions





Introduction

I One possible mode of operation for DesDyni is to use LIDARshots in a region in combination with the contiguous mapsproduced by SAR to better estimate forest structureseverywhere.

I This talk explores one way of classifying the a largeobservation area and determining underlying forest height andbiomass characteristics from areas where both SAR andLIDAR are available to areas where only SAR is available.

I We’ve previously presented results from our proof of concept(IGARSS ’09) using only simulated data and a small sample ofreal data (IGARSS ’ 10) we now present a working novelmulti-step classification algorithm.





Approach & High level algorithm

I Use simulators to estimate OPTICAL, LIDAR and SARmeasurements from 3D forest descriptions

I Generate many pine and spruce tree stands with a variety ofcanopy heights and biomasses to generate a stand databse

I Co-register OPTICAL, SAR, and LIDAR measurements in asingle image

I Compare each image pixel to the database and find the mostsimilar database stand





BOREAS Southern Study Area

I The SSA is approximately11,700 square kilometerscentered on 53.87299◦ Nlatitude and 105.2875 ◦ Wlongitude.

I A a confluence of multi-modalremotely sensed data exists from1994 - 1996.





Boreas Southern Study Area





BOREAS Southern Study Area: SAR & LiDAR Coverage





Algorithm Overview





Fractal Tree Model

I Model developed in late 1990’s.

I Fractal pseudo-random trees.

I Use Lindenmayer System:string-rewriting rules are used togenerate realistic branchingstructures, with needles andleaves.

I Each species of tree has its ownset of rules so it looks realistic.

I Both coniferous and deciduoustrees can be modeled.





Fractal Forest Model

Forest Attributes:

Biomass

Tree Species

Tree Attributes:

Height

Crown Diameter

Height to live crown

Trunk Diameter





SSA Jack Pine Stand





SSA Black Spruce Stand





Optical Model

I Use measured reflectance values for each canopy constituent:branches, trunks, leaves, needles, ground.

I Fractal geometry used with Pov-Ray ray-tracing code togenerate realistic 7-channel optical dataset.

I Rays are traced for many bounces

I Sensor is placed far above the forest, looking down at a 45◦

angle.

I Values are averaged over one pixel to produce the simulateddata.





SAR Model

I Use Foldy’s approximation to obtain the mean field in avertically-layered approximation to the canopy.

I Coherent simulation of each scattering mechanism: directcrown, direct ground, trunk-ground, crown-ground,crown-ground-crown,

I Fully-polarimetric.I Use at L band (1.25GHz)I All simulations at 20, 45, and 80 degrees incidence angle, 100

looks.I Interpolated polynomial best fit to allow for incidence angle

flexibility.I Validated at L band with measured SAR data (from BOREAS

and Raco, MI).





LIDAR Model

I Divide volume of stand into cubes.

I Each cube analyzed for what fractionof light is intercepted by thevegetation (cylinders and disks).

I Use vertical rays to estimate numberof intersections per cube.

I Radiative Transfer from cube-to-cubeto produce time-trace of LIDAR signal.

I Horizontal Gaussian pulse weightingacross the stand, with a verticalGaussian as well to obtain verticalresolution.





Radiative Transfer for one cube

I Given power propagating from aboveand below: quantify how muchtransmitted and reflected in eachdirection.

I Update the power propagating to nextcubes.

I Can use measurements from literatureto determine value for %reflected forbranches: 10%.

I Transmission through open areas isassumed 100%.

I Leaf transmission is assumed to be50%.





Database Overview

I Generated 4707 jack pine stands

I Generated 4364 black sprucestands

I All stands had a minimum of 10types of trees and up to 2000tree instances in an area of 625m2





Digital Elevation Model

I The BOREAS project generateda DEM in the 8th hydrologicalproject with 100m resolution

I A higher resolution DEM wasrequired for accurateorthorectification of the AirSARimages

I We created a 1315km by1390km km DEM byreprojecting and mosaicingnumerous DEMs from CDED.





SAR: AirSAR

I Numerous AirSAR images existin the Boreas SSA

I For this study, we selected twohigh resolution images with6.66m range resolution and9.26m azimuth resolution

I These images were orthorectifiedusing a DEM from CDED to asub-pixel accuracy of 6m





LiDAR: Scanning Lidar Imager of Canopies by EchoRecovery (SLICER)

I 37 Slicer flight paths wereconducted in the BOREASstudy areas in July 1996 yieldinga total of 834,277 LiDARwaveforms

I For each measurement, weextracted the power at canopyheight and the power ratiobetween the canopy height andthe ground return





LiDAR: SLICER

I Based on the location of eachmeasurement, a weightedaverage for both parameters wasderived for each





Optical: LandSAT7

I We used level 2T orthorectifiedLandsat data acquired in July1994

I Images were atmosphericallycorrected, cleaned of clouds andcloud shadows and reprojectedinto a single mosaic





Ground Truth

I Three data products from the BOREAS project were used asground truth for this study:

I Forest Species (Jack Pine or Black Spruce)

I Forest Biomass

I Forest Canopy Height

I Each ground truth data product was reprojected to 10mresolution cells (as needed)





Ground Truth - Tree Representations

I Stem mapped measurementswere recorded in the Jack Pinestands as well as the BlackSpruce stand.

I Using these measurements, wehave developed allometricequations to generate tree agiven species’ height to livecrown and diameter at breastheight as a function of thedesired tree height.

DBHjp = 0.0066h3 − 0.1404h2 +1.8672h − 1.9917

CHgtjp = 0.0001h4 − 0.0001h3 −0.0205h2 + 0.4788h − 0.7479

DBHbs =−0.0073h3 + 0.1708h2 + 0.2413h

CHgtbs =−0.0531h2 + 1.452h − 1.6152

R2 = 0.9564, 0.8555, 0.9442, 0.7133





Algorithm Overview





Level 0 Classification: Supervised Maximum LikilhoodClassification

I A simple two class classificationscheme was used: Trees andother.





Level 1 Classification: Database Comparison

I Each pixel containing AirSAR, SLICER, and LandSAT data aswell as ground truth data was examined

I Real remotely sensed values were compared to the 9000+simulated stands in our database

I The stand that most likely resembled the pixel underexamination was selected and that stand’s biomass and meancanopy height were assigned to the pixel





Error Function Measure

The error used is the weighted RMS error over the features:

1. 1.1 LIDAR mean power1.2 LIDAR peak power / LIDAR ground power1.3 SAR VV1.4 SAR HH

2. Optical Ch. 6

3. Optical NDVI

4. SAR VH

5. SAR VVHH





Introduction to Results

I Compare previous proof of concept to this study.

I Note that the proof of concept additionally used C-band SARand IfSAR

I Note that the proof of concept used the same forward modelsto generate our database as well as for the inversion andclassification.





Proof of Concept Results: Height





Proof of Concept Results: Biomass





Classification Results

I Classified 9071 pixels

I Species retrieval was 76.94% accurate.

I Height retrieval was 50.48% accurate with an RMS error of5.3m (to 7.3m).

I Biomass retrieval was 51.38% accurate with an RMS error of155.53 Ton/Ha.





Classification Results (2)

I If we know the target canopy will be small, under 13m, we canachieve even better results:

I Species retrieval was 76.94% accurate.

I Height retrieval was 67.16% accurate with an RMS error of4.37m.

I Biomass retrieval was 50.03% accurate with an RMS error of106.3 Ton/Ha.





Conclusions and Future Work

I We coregistered remotely sensed data from three differentsensors collected over a two year period.

I We generated a database with over 9,000 stands thatresemble those found in the BOREAS SSA.

I We created and implemented a multistep classification processwhich correctly identified the predominant tree species andwas over 50% accurate in identifying the canopy height andbiomass

I Future work includes introducing a recursive element to theL1 classification

I Future work includes the introduction of a multi-step errorfunction (used to select the most similar database stand)


benson_we3t051.pdf

Technology

optical data

sar lidar coveragebenson

sar andlidar

lidar measurements

globalclimate estimation

simulated data

d forest descriptionsgenerate

sensed data sets6