Mapping change in live and dead forest biomass

Mapping change in live and dead forest biomass with Landsat time-series, remeasured plots, and nearest-neighbor imputation Janet Ohmann 1, Matt Gregory 2, Heather Roberts 2, Robert Kennedy 3, Zhiqiang Yang 2, Justin Braaten 2, Scott Powell 4, Warren Cohen 1, Van Kane 5, and Jim Lutz 5 1 2 Department of Forest Ecosystems and Society, Oregon State University, Corvallis, OR 3 4 Vegetation Monitoring and Remote Sensing Team (VMa. RS) Resource Monitoring and Assessment Program (RMA) PNW Research Station, USFS, Corvallis, OR Department of Geography and Environment, Boston University, Boston, MA Department of Land Resources and Environmental Sciences, Montana State University, Bozeman, MT 5 College of the Environment, University of Washington, Seattle, WA Funded by: USDA-NIFA, USDA-FS Region 6 (Northwest Forest Plan)

Needs for regional vegetation information • Complexity and scope of forest policy issues (climate change, sustainability, etc. ) pushing technology to provide vegetation information that. . . – Supports integrated landscape analyses of multiple forest values – Consistent over large, multi-ownership regions (“all lands”) – Spatially explicit (mapped) – Detailed attributes of forest composition and structure – Latest challenge: trend information (monitoring)

Goal: develop a system for monitoring forest change over large regions • Can we marry two current technologies to better meet needs? – Land. Trendr to map change from Landsat time-series (trends) • – Nearest-neighbor imputation (detailed forest attributes) Illustrate methods with example forest biomass

Objective and strategies (nearest-neighbor component) • • Develop methods for integrating field plot and satellite data to monitor yearly changes in biomass – and other forest attributes (e. g. , older forest) Why nearest-neighbor imputation? – Multiple attributes in one map, serves many applications – Difficult-to-predict attributes swept along (e. g. species, CWD) – When k=1, maintain covariation among multiple attributes in map units, range of variability over study region

Gradient Nearest Neighbor Imputation (GNN)

The imputation component of GNN gradient space (1) conduct gradient analysis of plot data CCA Axis 2 (e. g. , climate) (3) find nearestneighbor plot(s) in gradient space (k=1, Euclidean distance in n-dimensional gradient space, with n axes weighted by explanatory power (eigenvalues)) geographic space (2) calculate axis scores of pixel from mapped data layers CCA Axis 1 (e. g. , elevation, wetness) (4) impute nearest neighbor’s value to pixel

Landsat Detection of Trends in Disturbance and Recovery (Land. Trendr)* • • Temporal normalization and segmentation at pixel level Minimizes noise from sun angle, phenology Segments describe sequences of disturbance, regrowth Normalized imagery for multiple years for GNN modeling *Kennedy et al. (2010), Rem. Sens. Env.

Matching plots to Land. Trendr imagery for GNN modeling • Unbalanced in time and space, 1 to 3 per location • Match plots to imagery of same year • Single gradient model from all plots, applied to each imagery year • Imagery is only change – assumes normalization • Response vars. : basal area by species + size-class 30 -m Landsat pixels 2 1 4 Yearly matching of plot and spectral data 3 FIA Annual plot

Preliminary results: Oregon and California Cascades region • 4, 812, 550 ha. (48, 126 sq. km. ) • n = 6, 207 plots

Dominant gradients in forest composition and structure (axis scores from CCA) % variation explained Axis 3 (blue): summer moisture stress, TC brightness Axis 1 (red): temperature, elevation, TC greenness, TC wetness Climate 8. 5 Landsat 4. 5 Topography 4. 2 Location 3. 7 Axis 2 (green): latitude, seasonality of precipitation

Biomass estimation: component ratio method (CRM) • • • National FIA approach (Heath et al. 2009) Component ratios from Jenkins et al. (2003) Bole volume equations from FIA-PNW: – Live trees (≥ 3 cm): 47 equations (species, DBH, height, location) – Snags (≥ 12 cm): Kozak’s taper equations • – Coarse woody debris (CWD) (≥ 25 cm): Smallian’s formula (small- and large-end diam. ) or Huber’s (intercept diam. only). Live (83%) > CWD (11%) > snag (6%) Jenkins components

Local- (plot-) level GNN map accuracy • • Cross-validation applies to all map dates, not spatial change Accuracy varies among attributes: – Best for live trees, density measures; poor for snags and CWD • GNN maps (k=1, Landsat) are inherently noisy, sensitive to slight spectral shifts, exacerbated by map differencing Live ≥ 3 cm Predicted (kg/ha) n. RMSE: 0. 54 r: 0. 78 r 2: 0. 60 Snags ≥ 12. 5 cm n. RMSE: 1. 59 r: 0. 43 r 2: 0. 21 CWD ≥ 25 cm n. RMSE: 1. 61 r: 0. 30 r 2: -0. 08

Regional accuracy (area distributions) -- 2005 Live ≥ 3 cm (kg/ha) Snags ≥ 12 cm (kg/ha) • CWD ≥ 25 cm (kg/ha) • • GNN single year (2005) vs. FIA Annual sample (2001 -2010) GNN (k=1) maintains distributions in plot sample, even for attributes poorly predicted at local scale Map has no “unsampled” area

Agreement Coefficient (Ji and Gallo 2006) • • • Compares GNN to FIA Annual plots (n=1, 560) over a range of scales ACUNS poor at plot scale, improves with scale ACSYS is good at all scales Live above-ground biomass (trees ≥ 3) (Mg/ha) ACUNS: -0. 39 ACSYS: 0. 92 Plots ACUNS: 0. 66 ACSYS: 0. 96 ACUNS: 0. 92 ACSYS: 0. 97 10 -km hexagons (8, 660 ha) 30 -km hexagons (78, 100 ha)

Mapping live biomass change, 1990 -2010 10 km - - 10 miles -- Land. Trendr 1990 B-G-W 2010 B-G-W Disturbance GNN 1990 (mg/ha) 2010 (mg/ha) Change (mg/ha)

Land ownership and allocation in the Oregon and California Cascades movie location 90 km

Total live above-ground biomass (trees ≥ 3 cm) (1990 – 2010) Nonforest Mg/ha (metric tons) 0 – 54. 6 – 125. 7 – 202. 4 – 281. 8 – 373. 0 – 475. 8 – 581. 0 – 724. 0 – 884. 1 – 1, 374. 4 90 km

Biomass of snags ≥ 12. 5 cm (1990 – 2010) Nonforest Mg/ha (metric tons) 0 – 6. 2 – 15. 6 – 25. 9 – 37. 7 – 54. 2 – 76. 0 – 99. 6 – 171. 1 – 253. 3 – 378. 1 90 km

Disturbance and live biomass change (%), 1990 -2010 • Net loss in harvested landscapes (nonfederal) • Net loss in areas of natural disturbance (fire, I&D) • Net gain in areas of regrowth (federal reserves) Land. Trendr % disturbed GNN live % change

Disturbance and snag biomass change (%), 1990 -2010 • Net loss in harvested landscapes (nonfederal) • Net gain in areas of natural disturbance (fire, I&D) Land. Trendr % disturbed GNN snag % change

Disturbance and CWD biomass change (%), 1990 -2010 • Disturbance effects are more subtle, at least in the short term Land. Trendr % disturbed GNN CWD % change

Disturbance and biomass change (live, snag, down) • All disturbances are not created equal: – Harvest results in net biomass loss (live and snag) • – Natural disturbances result in flux between live and dead pools Components of structural diversity don’t vary in tandem – Disturbance ≠ biomass (or habitat) loss • – Live biomass loss may coincide with gain in diverse early seral forest Patterns of change are complex in both time and space – Stay tuned! Live Snag CWD

Lessons learned: multi-temporal GNN for monitoring • Only feasible with “temporally normalized” imagery – Reliability rests on assumption of normalization, particularly in older, undisturbed, stable forest – Technical challenges are substantial, especially in “production” • Large error in mapped change at pixel scale, but reasonable trends emerge at coarser grains