International Journal of Applied Earth Observation and Geoinformation
Effect of topographic correction on forest change detection using spectral trend analysis of Landsat pixel-based composites
Introduction
The open release of the Landsat image archive in 2008 (Woodcock et al., 2008) has resulted in an increase in the number of spatial and temporal processing and analytical methods (Banskota et al., 2014) associated with large-area analyses from remotely sensed data (Wulder and Coops, 2014). No longer impeded by per-image costs and further enabled by provision of analysis-ready products that have quality geometric co-regisration and calibrated spectral values, users are increasingly empowered (Hansen and Loveland, 2012, Wulder and Coops, 2014), especially for time series investigations (Kennedy et al., 2010). Multiple images with analogous conditions can be combined to create spatially exhaustive image composites based upon individual pixels (Griffiths et al., 2013a, Griffiths et al., 2013b, Roy et al., 2010, White et al., 2014) rather than a need for cloud-free scenes, with automated algorithms utilized to remove cloud, shadows, and other atmospheric effects (e.g., Zhu and Woodcock, 2012). Access to dense time series and an exhaustive spatial coverage combined with a level of detail that is informative of human and management activities on terrestrial ecosystems has been afforded by Landsat (Wulder et al., 2008). Change detection, with a particular emphasis on forest change, has been a particularly active research area (e.g., Kennedy et al., 2010, Huang et al., 2010), with opportunities to not only capture stand replacing change, but also to relate more subtle changes including magnitude, duration, and preceding and following land cover conditions (Frazier et al., 2014, Hermosilla et al., 2015a).
Currently, analysis-ready Landsat Level 1T (standard terrain correction) products include systematic radiometric and geometric corrections, and include the use of a digital elevation model (DEM) for topographic accuracy (USGS, 2013), noting that there is no standard topographic correction applied for terrain shadows. Topographic correction is an image enhancement used to minimize the impact of shadows or varying illumination caused by terrain (Richter, 1998), the effects of which are not well explored regarding Landsat time series imagery and forest change (Banskota et al., 2014). Pixel-based compositing techniques produce images with a range of pixel acquisition dates, resulting in the possibility of terrain shading differences due to a variation in solar azimuth and zenith. Certain image processing activities can be impacted by differences in surface reflectance values that are not indicative of actual surface condition differences, but rather a manifestation of the physical location (see Tan et al., 2013). Varying terrain shadows throughout the year, as may be the case in a time series based on composites, may disrupt spectral trends, potentially causing errors in change detection results (Banskota et al., 2014). However, studies exploring forest change detection and pixel-based composites have not included topographic corrections (e.g., Griffiths et al., 2013a, Griffiths et al., 2013b, Hermosilla et al., 2015a), and a study that investigated the effects of topographic correction on a land cover classification of pixel-based composites found only a small increase in classification accuracy (Vanonckelen et al., 2015).
The current capacity for processing of large numbers of images for both large area and dense time series is evident (e.g., Griffiths et al., 2012, Griffiths et al., 2013a, Griffiths et al., 2013b, Latifovic and Pouliot, 2014, Hermosilla et al., 2015a, Senf et al., 2015) with the resulting composites supporting monitoring and reporting programs (White et al., 2014). To ensure the quality of these outcomes, additional investigation to determine the possible impacts of spectral differences attributable to topography on land cover and change applications is required. Topographic correction routines remain time consuming, require access to adequate digital terrain models, and can be implemented in a variety of ways (e.g., Soenen et al., 2005, Richter et al., 2009), with no consensus on the most appropriate approaches to follow, some of which may be land cover or location dependent. Knowledge of whether or not to include a geometric topographic correction in an image compositing and subsequent analysis workflow is an important consideration for resultant product quality as well as for time and data management. In this research we explore the effects of topographic correction on change detection outcomes using spectral trends computed from Landsat imagery. Specifically, this research focuses on the differences in the quantity of change detected in forested areas as well as the variation in this quantity across different topographic positions when using uncorrected composites, and composites from two different topographic correction approaches, one correcting each pixel considering its actual acquisition day of year (DOY), and the other considering a single date for all pixels.
Section snippets
Study area
The study area is 56,000 km2 and located in southwest British Columbia, Canada. The location is well-suited to examine the effects of mountainous terrain on spectral trends as it includes both the Coast Mountain Range and the Cascade Mountain Range with an elevation range from 0 m to 3236 m (mean: 1401 m, standard deviation: 544 m), a mean slope of 20°, and various complex terrain features, such as valleys, ridges, and avalanche chutes (Fig. 1). This study area includes highly productive temperate
Methods overview
First, three sets of annual Landsat pixel image composites from 1984 to 2012 were created using: (i) no topographic correction, (ii) a day of year topographic correction (hereafter TCDOY) wherein each pixel in the composites are corrected according to their actual acquisition date, and (iii) an August 1st topographic correction (hereafter TCAug1) wherein each pixel is corrected as though it were acquired on August 1st. Second, change detection using spectral trend analysis was applied to each
Results
Overall, both the TCDOY and the TCAug1 change detection outcomes had less total area changed than NC outcome (Fig. 3). NC produced the largest area detected as change, followed by TCAug1 and TCDOY. For TCDOY correction, there is less area identified as change than for the non-corrected composite in 21 of the 27 analyzed years (Fig. 3). For the TCAug1 correction, the correction process detects less area as change in 23 of the 27 years (Fig. 3). The area detected as change each year ranges from
Discussion
When undertaking change detection applications, images would ideally be from similar dates across years (often called anniversary dates). Given the satellite revisit cycles, image availability and quality, there has been a tolerance around the anniversary date as a target, not an absolute requirement. With pixel-based composites, in contrast to scene-based change analyses, there is an opportunity for greater variability in dates between years due to the pixel-based compositing approach. That
Conclusions
In this research we analyze the effect of topographic correction on forest change predicted from time series of Landsat pixel-based image composites. To do this, we topographically corrected the image composites using the sun-canopy-sensor correction algorithm. The spectral trends of the composites were analyzed to detect non-stand-replacing changes. We assessed the effects of the algorithm using two different correction dates: August 1st, in which each pixel was corrected as though it had been
Acknowledgements
Geordie Hobart of the Canadian Forest Service is acknowledged for preparation of the annual pixel-based composites used in this research. This research was undertaken as part of the “National Terrestrial Ecosystem Monitoring System (NTEMS): Timely and detailed national cross-sector monitoring for Canada” project jointly funded by the Canadian Space Agency (CSA) Government Related Initiatives Program (GRIP) and the Canadian Forest Service (CFS) of Natural Resources Canada. Additional support was
References (37)
- et al.
A sun-crown-sensor model and adapted C-correction logic for topographic correction of high resolution forest imagery
ISPRS J. Photogramm. Remote Sens.
(2014) - et al.
Characterization of aboveground biomass in an unmanaged boreal forest using Landsat temporal segmentation metrics
ISPRS J. Photogramm. Remote Sens.
(2014) - et al.
Using annual time-series of Landsat images to assess the effects of forest restitution in post-socialist Romania
Remote Sens. Environ.
(2012) - et al.
Topographic normalization of Landsat TM images of forest based on subpixel sun-canopy-sensor
Remote Sens. Environ
(1998) - et al.
A review of large area monitoring of land cover change using Landsat data
Remote Sens. Environ.
(2012) - et al.
An integrated Landsat time series protocol for change detection and generation of annual gap-free surface reflectance composites
Remote Sens. Environ.
(2015) - et al.
An automated approach for reconstructing recent forest disturbance history using dense Landsat time series stacks
Remote Sens. Environ.
(2010) - et al.
Detecting trends in forest disturbance and recovery using yearly Landsat time series: 1. LandTrendr — Temporal segmentation algorithms
Remote Sens. Environ.
(2010) - et al.
Forest cover trends from time series Landsat data for the Australian continent
Int. J. Appl. Earth Obs. Geoinf.
(2013) - et al.
The development of 1901–2001 historical monthly climate models for Canada and the United States
Agric. For.
(2006)