Oregon Department of Forestry - PowerPoint Presentation

Slide1

Oregon Department of ForestryForest Inventory Systems and Lidar

Operationalizing Lidar in Forest Inventory

Tod Haren

1/25/2016 – Olympia, WASlide2

IntroductionsOverview of ODF (State Forests)Inventory Tool Chain

Stand Level Inventory

Data ManagementStand Level ImputationLidar ProcessingLandsat ProcessingRandomForests ImputationWhat is being inventoriedWho is being servedWhat outputs are producedHow are species handledHow are outputs integrated with existing systemsHighest priorities for improvementSlide3

ODF Staff PresentJeff Firman – Forest Inventory SpecialistMike Wilson – GIS and Information Specialist

Sephe Fox –

GIS and Information SpecialistJosh Clark – Modeling Lead, Forest PlanningTod Haren – Forest Resource AnalystSlide4

ODF – State Forests800,000+ acres under management

3% of Oregon’s forested land

Operations funded by timber sale revenueNine management districtsSix state forestsSlide5

Common School LandEvery 16th and 36th section granted at statehood

Forested parcels managed by ODF under agreement with DSL

Net timber sale revenue contributed to the Common School FundODF is reimbursed for management costsElliott State Forests~84,000 acres resulting from a series of land tradesSlide6

Board of Forestry LandAcquired from counties following tax foreclosure

Tillamook burn

1933, 1939, 1945, 1951350,000+ acresPlanting and rehab continued into the 1970’sManaged for “Greatest Permanent Value”63.75% of timber sale revenue distributed to the county in which harvest occursSlide7

Stand Level InventoryFBRI - Forest Projection System (v6.4+)

Designed for traditional double sampling

Stands delineated on dominant vegetationStratification using the FPS veg_lbl methodDominant species; size class; stocking levelInitiated in ~2000 to support reporting needs anticipated for the then new forest management plans.Additional fields and tables addedMS Access front-end developed with significant VBA codeIn-house developed field data collection software using DataPlusSlide8

Stand Level InventoryOriginally promoted decentralized i

nventory to give more local control

Local control quickly became burdensomeAttrition, burnout, inadequate training, etc.Standards were not being enforcedReporting and analysis was inconsistent and cumbersomeSalem is now taking a more active rollMost annual updates funneled through the inventory specialistMore communication and coordinationMigration to a central database - SQL Server and ArcGIS/SDEImproved access to current and historic data

Ready access to pertinent information for all staff

Better systems integration

Annual updates and “ROOTS” still require individual Access databasesSlide9

Stand Level InventoryInitial goal was to maintain 50% of stands

as recently cruised

In 2008 Tillamook undertook a significant re-typingMany cruised stands were split, invalidating plot designRetained plots are questionable due to reconfiguration of standsSampling curtailed due to budget cuts in 2009Reinitiated in 2015NW Planning area inventory status (2014)Slide10

Stand Level InventoryProportional sample allocation by strata

Some input from field based on operational priorities

Plots located along lines to represent typical cover, crossing topographic features and elevation gradientTypically 16-24 plots per standSlide11

Stand Level InventoryNested plot design

Large tree variable radius

Small trees, snags, understory fixed radiusDown wood line transectSpecies, DBH, Damage for all treesHeights subsampled by speciesSite trees selected from dominant & codominant canopyNo radial incrementNo upper stem measurement (form)Plots not geo-located, mapped points are preliminarySlide12

Stand Level InventoryConverted from strata expansion to imputation in 2008Better representation of variability across the landscape

Forester “best guess” assignment of cruised to non-cruised stands

No formal validation processTillamook – 6000+ stands, only 20% cruised (fewer in 2008)Less confidence in imputation as inventory agedMaintenance became burdensomeSlide13

Tillamook ImputationRoughly 70%

lidar

coverage by 2011Central swath covered in 2012NE corner partially flown in 2015Slide14

Tillamook Imputation2012 Cooperative agreement with RMRS, Moscow, ID to develop imputation methods using

lidar

and landsatHudak, Andrew T.; Haren, A. Tod; Crookston, Nicholas L.; Liebermann, Robert J.; Ohmann, Janet L. 2014. Imputing forest structure attributes from stand inventory and remotely sensed data in western Oregon, USA. Forest Science. 60(2): 253-269.All stands projected to each of the lidar flight yearsTested multiple imputation methods, MSN, GNN, RandomForestsTested imputation of stand signatures onto pixelsSelected RandomForests stand level imputationDependent variables included: TPA; BAA; SDI; CCF; HT; QMD; Tot VPA;

Merch

VPA; Tot Carbon

Nine independent variables selected based on scaled importance

Lidar: Top

Ht

; Return density in vertical strata; 25

th

pct

ht

Landsat: Brightness; Greenness; Wetness; Topographic variables

Evaluated plot level imputation, but results were very poor due to lack of geo-locationSlide15

Tillamook ImputationLidar imputation:

Observed vs. Imputed

Stand level vs. Pixel levelA&B – Stand to standC&D – Stand to pixelLandsat pixel level is even more skewed toward meanSlide16

Tillamook ImputationComparison with 402 stands cruised in 2010

Plot compares difference between imputed and cruised VPA

Landsat and lidar compare observed values with nearest neighbor402 stands cruised in 2009 compared with the previous imputation assignmentsStrata Exp compares against strata averageSlide17

Tillamook Imputation – 2014 UpdateAdditional cruised stands

Central swath

lidar data, ~90% total coverageAll stands grown to 2012Included additional dependent variablesReduced the precision for any one variableBetter(?) representation of key structure variablesIntegrated landsat and lidar, top 40 variables in final modelLidar variables most important for basic stand attributesLandsat variable became important as species and structure attributes were added.Slide18

Tillamook Imputation – 2014 Update

2014 Observed vs Imputed

TotalCubic VPA; Scribner VPA; TPA >=8”; SDI >=8”; Top Ht; Top QMD; BAA Std DevConiferScribner VPA; SDI >=8”; TPA >= 8”DF; WH; RAScribner VPA; Top HtSlide19

Imputation ToolsPython

PyODBC

- SQL Server and MS Access data managementLiblas - Catalog lidar (laz) datasets OGR - Tiling lidar catalogs, post processing pixel level dataParallel Python - Asynchronous execution of Fusion callsSQLite - Aggregation of tiled Fusion metrics

Need to evaluate the

SciKit

-Learn package

RandomForest

implementation as well as many other machine learning algorithms

R

YaImpute

-

RandomForest

model development and evaluation

GGPlot2

-

PlottingSlide20

ODF Forest InventoryWho is being served

Field foresters – Annual operational planning; T&E assessment

DSL & Counties – Annual reportingBOF, Stakeholders, Managers – Long-range planning; harvest modeling; growth and yield analysisWhat outputs are producedStatic reports – Cruise stats; stand tablesGeoPlanner – GIS overlays, SQL stored procedures generate dynamic stand summaries and prescription analysisYield tables – SQLite database integrated with Patchworks harvest scheduling software.Slide21

ODF Forest InventoryHow are species handled

Stand level imputation is single nearest neighbor

Target stand assumes all plot attributes of the source standSpecies level stand attributes are included in the imputation modelLandsat multispectral variables, tasseled cap, band rations, veg. indices improved species level response in the imputation modelIntegrating with existing systemsImputation assignments stored in the [ADMIN] database table and used as a foreign key to the cruise tablesSlide22

ODF Forest InventoryHighest priorities for improvementEvaluation of small area estimation and pixel based methods

What efficiencies can be gained – Fewer field samples, lower cost

What is the potential for increased information (precision and accuracy)Need more working examples of alternative inventory methodsWhat are the costs; what is to be gainedHow do the results compare with stand based samplingExamples of integration with growth and yield, eg. tree list generation

If we ultimately stick with stand based inventory

Does our sampling design adequately represent within stand variation

Could we use a multiple neighbor approach

What about separate imputation models for subsets of stand attributes

Data management becomes tricky with anything beyond single neighbor