Ensemble Integration of Forest Disturbance Maps for the Landscape Change Monitoring System (LCMS)

Friday, 19 December 2014: 9:45 AM
Warren B. Cohen1, Sean P Healey2, Zhiqiang Yang3, Zhe Zhu4, Curtis E Woodcock4, Robert E Kennedy4, Chengquan Huang5, Daniel Steinwand6, James E Vogelmann7, Stephen V. Stehman8 and Thomas R Loveland9, (1)US Forest Service Corvallis, Corvallis, OR, United States, (2)Rocky Mountain Research Statio, Ogden, UT, United States, (3)Oregon State University, Corvallis, OR, United States, (4)Boston University, Boston, MA, United States, (5)University of Maryland College Park, Geographical Sciences, College Park, MD, United States, (6)USGS, EROS Data Center, Sioux Falls, SD, United States, (7)USGS EROS Center, Sioux Falls, SD, United States, (8)SUNY College of Environmental Science and Forestry, Syracuse, NY, United States, (9)USGS EROS Data Ctr, Sioux Falls, SD, United States
The recent convergence of free, high quality Landsat data and acceleration in the development of dense Landsat time series algorithms has spawned a nascent interagency effort known as the Landscape Change Monitoring System (LCMS). LCMS is being designed to map historic land cover changes associated with all major disturbance agents and land cover types in the US. Currently, five existing algorithms are being evaluated for inclusion in LCMS. The priorities of these five algorithms overlap to some degree, but each has its own strengths. This has led to the adoption of a novel approach, within LCMS, to integrate the map outputs (i.e., base learners) from these change detection algorithms using empirical ensemble models. Training data are derived from independent datasets representing disturbances such as: harvest, fire, insects, wind, and land use change. Ensemble modeling is expected to produce significant increases in predictive accuracy relative to the results of the individual base learners.

The non-parametric models used in LCMS also provide a framework for matching output ensemble maps to independent sample-based statistical estimates of disturbance area. Multiple decision trees “vote” on class assignment, and it is possible to manipulate vote thresholds to ensure that ensemble maps reflect areas of disturbance derived from sources such as national-scale ground or image-based inventories.

This talk will focus on results of the first ensemble integration of the base learners for six Landsat scenes distributed across the US. We will present an assessment of base learner performance across different types of disturbance against an independently derived, sample-based disturbance dataset (derived from the TimeSync Landsat time series visualization tool). The goal is to understand the contributions of each base learner to the quality of the ensemble map products. We will also demonstrate how the ensemble map products can be manipulated to match sample-based annual TimeSync estimates of disturbance area.