An LTSS is defined as a temporal sequence of Landsat images acquired at a nominal temporal interval for an area defined by a path/row tile of the World Reference System (WRS)� These images should have an IRU quality, which is defined as follows: They are the best available images, and they have high levels of geolocation accuracy and radiometric integrity� The temporal frequency of observations in an LTSS is driven by data availability and the temporal characteristics of the changes to be detected� For forest change analysis, annual or seasonal observations are desirable, but if such frequent observations are not available, LTSS with biennial or longer temporal intervals may also be used, especially in high-latitude regions where trees grow slowly� However, due to limited data availability and cloud contamination, the actual temporal intervals between consecutive acquisitions in an LTSS can be different from the nominal interval of that LTSS (example acquisition dates of some LTSS can be found in the study by Huang, Goward, Schleeweis et al� 2009;
Table 14�1)�
The process for developing an LTSS comprises an image selection protocol, automated high-level preprocessing algorithms, and IRU quality-verification procedures (Figure 14�1)�
The high-level preprocessing algorithms include updated radiometric calibration for Land sat- 5 images, atmospheric adjustment to surface reflectance, precision registration, and
orthorectification� Here, the term “high level” is used to differentiate these algorithms from the standard correction algorithms for Landsat images (Landsat Project Science Office 2000)�
These high-level preprocessing algorithms have been implemented as fully automated routines in the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) to facilitate batch job processing�
14.2.1 Image Selection
The main purpose of image selection is to identify high-quality Landsat acquisitions that are needed to constitute an LTSS� Because Thematic Mapper (TM) and Enhanced Thematic Mapper+ (ETM+; except the panchromatic band) images have almost identical spatial and spectral characteristics, they are used interchangeably in the LTSS, as in many other land- cover change analyses (e�g�, Vicente-Serrano, Pérez-Cabello, and Lasanta 2008; Lo and Yang 2002)� The following two issues need to be considered in image selection:
1� A selected image must be acquired during the leaf-on season� Images acquired outside this temporal window are generally not suitable for forest change anal- ysis, because leaf-off deciduous forests can be spectrally confused with dis- turbed forest land� For each WRS path/row, the leaf-on season can be defined based on vegeta tion phenology derived using Moderate Resolution Imaging Spectroradiometer (MODIS) and Advanced Very High Resolution Radiometer (AVHRR) measurements (Schwartz, Reed, and White 2002; Zhang et al� 2003), which includes June to mid-September for most areas in the conterminous United States� This criterion can be relaxed to include May and October for the southern United States�
2� In order to maximize the proportion of usable pixels within each selected image, it should have minimal or no quality problems arising from instrument errors or from cloud and shadow contamination�
14.2.2 updated radiometric Calibration
Radiometric calibration is part of the U�S� Geological Survey (USGS) process for producing standard level-1 Landsat imagery (Landsat Project Science Office 2000)� For Landsat-7 images, the ETM+ sensor has been monitored since its launch (Markham et al� 2004)� As such, the conversion of ETM+ level-1 radiometry to at-sensor radiance is a simple mat- ter of applying the rescaling gains and biases from the ETM+ header file to the imag- ery� However, for Landsat-5 images, there have been many revisions to the calibration
Table 14.1
Standard Deviation Values (Reflectance in Percentage) Used in Equations 14�4 and 14�5
Band 1 Band 2 Band 3 Band 4 Band 5 Band 7
0�80 0�582 0�617 3�575 1�214 0�768
Source: Huang, C� et al� Remote Sens Environ, 114, 1, 2010� With permission�
Note: These values were the average of those derived using images acquired in different years from different places�
coefficients (Chander et al� 2004; Chander and Markham 2003; Markham and Barker 1986;
Markham et al� 2004)� Prior to May 2003, internal calibration (IC) lamps measurements were used to determine the gain coefficient� During the 1990s, it became increasingly appar- ent that variations in the IC-derived gain values reflected a combination of real changes in sensor calibration (i�e�, detector sensitivity and filter properties) and changes in the IC lamps themselves�
From May 2003 to April 2007, following the Landsat-7/Landsat-5 underfly experi- ment in April 1999, a cross-calibration between ETM+ and TM was established (Chander et al� 2004; Teillet et al� 2004), which was used to determine the gain value for Landsat-5 images� After April 2007, further investigations using invariant ground targets in North Africa suggested that the initial lookup table (LUT) had an error for the first part of the Landsat-5 history (about 1985–1992)� Instead, records of at-sensor radiance from these sites suggested a more gradual decay in gain throughout the mission life� As such, Landst-5 imagery processed after April 2007 used a revised LUT (Chander et al� 2004)�
The production date of a Landsat-5 image needs to be used to determine the calibration that was originally applied to that image, which can then be “undone” by applying the reciprocal of the gain, and then the most recent LUT can be applied� Conversion to top-of-atmosphere (TOA) reflectance is then performed for the reflective bands using the scene-specific solar geometry, and sensor-specific bandpasses convolved with the CHKUR exoatmospheric irradiances from MODTRAN-4 (Landsat Project Science Office 2000;
Markham and Barker 1986)�
For the thermal band, the raw digital number is converted to TOA (apparent) tempera- ture using the standard approach provided by Markham and Barker (1986) for TM images and by the Landsat Project Science Office (2000) for ETM+ images� For Landsat-5 images, however, a radiance correction to the calibration published in late 2007 needs to be consid- ered (Barsi et al� 2007)�
14.2.3 atmospheric adjustment to Surface reflectance
The LEDAPS atmospheric adjustment algorithm was designed to calculate surface reflectance by compensating for atmospheric scattering and absorption effects on the TOA reflectance (Masek et al� 2006)� The basic assumptions of this algorithm are that the target is Lambertian and infinite and the gaseous absorption and particle scattering in the atmo- sphere can be decoupled�
Developed based on a similar method used for MODIS and AVHRR (Vermote, El Saleous, and Justice 2002), this approach uses the 6S (second simulation of a satellite signal in the solar spectrum) radiative transfer code to compute the transmission, intrinsic reflectance, and spherical albedo for relevant atmospheric constituents, including gases, ozone, water vapor, and aerosols (Vermote et al� 1997)�
Ozone concentration was derived from the Total Ozone Mapping Spectrometer (TOMS) aboard the Nimbus-7, Meteor-3, and Earth Probe platforms as well as from the National Oceanic and Atmospheric Administration’s (NOAA) Tiros Operational Vertical Sounder ozone data when TOMS data was not available� Column water vapor was taken from NOAA National Centers for Environmental Prediction (NCEP) reanalysis data (available at http://dss�ucar�edu/datasets/ds090�0)� Digital topography (1-km GTOPO30) and NCEP surface pressure data were used to adjust Rayleigh scattering to local conditions� Aerosol optical thickness was directly derived from the Landsat image using the dark, dense veg- etation method of Kaufman et al� (1997)�
14.2.4 Precision registration and Orthorectification
Raw satellite images usually contain significant geometric distortions arising due to a range of sources, including platform- and instrument-related sources, as well as those due to the Earth’s curvature, rotation, and topography (Toutin 2003)� Beginning from late 2008, the USGS decided to distribute terrain-corrected (L1T) images as the standard Landsat imagery product� In general, no additional geometric correction is necessary for those L1T images because their geolocation errors are typically less than 1 pixel� Unfortunately, as of the writing of this chapter, this USGS standard has not been adopted by international ground-receiving stations� The standard systematic correction (L1G) products distributed by those stations can have geolocation errors of 500 m or more� This results primarily from uncontrolled orbital drift� Further analysis of image relief displacement as a result of topography has shown that at swath edges geolocation is in error by about 120 m per kilometer of elevation� Therefore, the LEDAPS precision registration and orthorectification algorithm are still needed for images obtained from international ground-receiving sta- tions� Details of this algorithm have been provided by Gao, Masek, and Wolfe (2009)�
14.2.5 landsat Time-Series Stacks Imagery-ready-to-use Quality Verification
Prior to its use in downstream applications, each developed LTSS needs to be verified to determine whether the processed images have geometric or radiometric artifacts, which can result from the following:
Unidentified quality problems with the input images
•
Unknown bugs that may exist in the LEDAPS preprocessing algorithms
•
Incorrect inputs regarding the geometry or radiometry of the concerned images
•
If artifacts are found in some LTSS images, they need to be fixed, or the images contami- nated by those artifacts need to be excluded from downstream change analysis to avoid resulting in spurious changes� The verification procedures include a quick visualization approach and a spectral–temporal profile method� Prior to verification, the images of each LTSS need to be clipped so that they have exactly the same spatial domain�
14.2.5.1 Image Clipping
Due to difficulties in orbital control, satellite orbits can shift slightly among repeat passes�
As a result, several images for a single WRS location are not necessarily congruent to the same geographic region (i�e�, they do not overlay on top of each other exactly)� Therefore, pixels near the edge of a WRS tile can have valid values on some dates but not on other dates� Temporal analysis of such incomplete observations is difficult� To avoid this problem, a common area mask is used to exclude such observations from being analyzed� For each LTSS, this mask is defined as the maximal geographic area where all image acquisitions have valid pixel values (except the missing data area caused by missing scan lines that may exist in certain acquisitions), and all images of the LTSS are clipped using this mask�
14.2.5.2 Visual Verification
Visual inspection is a simple yet effective method for verifying the quality of the LTSS images� By flipping the images from one date to another, an experienced image analyst can
quickly identify inconsistencies among the images, which are often indicators of geomet- ric or radiometric artifacts, and can gain first-hand knowledge of the change processes to be analyzed later� To facilitate quick visualization of the LTSS images, the clipped images are converted to the JPEG format (note that other visualization-ready formats can be used in the place of the JPEG format here), which are then assembled to create a movie loop�
For each LTSS, a single-stretching method is used during the conversion in order to create comparable color tones among the images� After testing with different stretching methods, it was found that a single-stretching method per band could produce satisfactory visual effects for images acquired in most areas consisting primarily of closed or near-closed canopy forests� The general linear stretching equation is
out_value in_value refl_min refl_max ref
= −
− l_min×255 (14�1)
where in_value and out_value are the input surface reflectance (%) and output stretched values, respectively� The single set of refl_min and refl_max applicable to most vegetated areas are given in Table 14�2�
In semiarid areas, some partially vegetated areas may appear saturated in bands 2 and 3 when stretched using Equation 14�1� To avoid this problem, the following nonlinear stretching method is used for those two bands for images acquired in such areas (band 4 is stretched using Equation 14�2):
Out value_ =13×In value_ −0 15� ×In value_ 2 (14�2) 14.2.5.3 Spectral–Temporal Profile
Spectral–temporal profiles can provide a more quantitative assessment of the radiomet- ric consistency among the images within each LTSS� Such profiles are created using the spectral values of targets that are considered relatively stable over time� Because the pri- mary use of the LTSS assembled here is forest change analysis, the “stable targets” in this context refer to conifer stands that do not have visual signs of being disturbed during the entire observing period of each LTSS� For each LTSS, a few examples of such stands are identified, and the average spectral values of those stands are calculated for each acquisi- tion date� The values for all dates are then plotted as a function of the acquisition date�
Figure 14�2 shows that the TOA reflectance values can vary greatly, mostly due to changing atmospheric conditions from year to year� Most of those variations are removed or greatly reduced by performing atmospheric corrections�
Table 14.2
Parameters Used in Equation 14�1 for Stretching the Surface Reflectance Images to Create JPEG Images for Use in the Movie Loop
Refl_min (%) Refl_max (%)
Band 2 0 15
Band 3 0 15
Band 4 2 50
Source: Huang, C� et al� Int J Digital Earth, 2, 3, 2009� With permission�
Note: Only standard false color images were created using the TM/ETM+ bands 4, 3, and 2 shown in red, green, and blue colors�