Abstract
Remote detection of forest disturbance remains a key area of interest for scientists and land managers. Subtle disturbances such as drought, disease, insect activity, and thinning harvests have a significant impact on carbon budgeting and forest productivity, but current change detection algorithms struggle to accurately identify them, especially over decadal timeframes. We introduce an algorithm called Edyn, which inputs a time series of residuals from harmonic regression into a control chart to signal low-magnitude, consistent deviations from the curve as disturbances. After signaling, Edyn retrains a new baseline curve. We compared Edyn with its parent algorithm (EWMACD—Exponentially Weighted Moving Average Change Detection) on over 3500 visually interpreted Landsat pixels from across the contiguous USA, with reference data for timing and type of disturbance. For disturbed forested pixels, Edyn had a mean per-pixel commission error of 31.1% and omission error of 70.0%, while commission and omission errors for EWMACD were 39.9% and 65.2%, respectively. Edyn had significantly less overall error than EWMACD (F1 = 0.19 versus F1 = 0.13). These patterns generally held for all of the reference data, including a direct comparison to other contemporary change detection algorithms, wherein Edyn and EWMACD were found to have lower omission error rates for a category of subtle changes over long periods.
Highlights
Remote sensing, with Landsat in particular, has long been used to detect forest disturbance at a wide variety of scales [1,2,3,4]
In this study we describe the differences between Edyn and Exponentially Weighted Moving Average Change Detection (EWMACD), compare the two algorithms in an agreement assessment on over 3500 forested pixels from across the contiguous United States (CONUS) using decades-long time spans
For reference data in our assessment, we used a collection of pixels interpreted by users of the TimeSync software [23], a web application hosted by Oregon State University. (TimeSync version 2.0 was used for generating the reference data.) These pixels were randomly sampled from a collection of 179 Landsat scenes covering a diverse subset of the CONUS ([13], Figure 1), with each scene hosting different forest types and disturbance regimes
Summary
With Landsat in particular, has long been used to detect forest disturbance at a wide variety of scales [1,2,3,4]. Cohen et al [13] found evidence of a more general forest decline, defined there as canopy loss not associated with or attributable to other common classes (in the case of that study: fire, harvest, wind, water, land use conversion, or debris), that has a wide extent and an increasingly sizable impact on forest productivity and carbon flow Given that these subtle changes can affect the structure and Forests 2017, 8, 304; doi:10.3390/f8090304 www.mdpi.com/journal/forests functioning of forests over large areas, there is a great need to identify them and other such withinclass changes (e.g., a thinned or drought-stressed forest) using remote sensing
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have