In remote sensing, both observational scale and measurement scale are prominent characteristics. Any particular imaging device is specifically designed to make observations of a defined extent and resolution. The scales range from high-resolution imagery with a sampling size of several centimeters to several kilometers and from local observations to global coverage. In the process of image acquisition in optical remote sensing, the two notions of scale are inextricably interconnected: an image of high resolution covers an area of small extend while lower resolution data cover a larger area. Especially the characteristic of varying resolution and the abundance of remote sensing data have fostered the exertion of scale studies using remote sensing data (Marceau, 1999). Thus, a multitude of research has been conducted in regard of scales and scaling under different aspects (for example Friedl et al.,
1995, Friedl 1997, Hay et al. 1997, Pax-Lenney and Woodcock, 1997, Pelgrum, 2000, Treitz and Howard, 2000, Tian et al. 2002, Garrigues et al. 2002, Ju et al. 2005)
The issue of spatial sampling resolution in remote sensing data is additionally linked to the aspect of temporal resolution. Data of highest resolution are mostly collected from airborne instrumentation and are limited to localized and infrequent campaigns. In contrast, most spaceborne imagers are designed to continuously collect data, yet the frequency of observation diminishes with resolution increase. These conditions are decisive when remote sensing data are to be operationally used. They are also important in the attempt to incorporate remote sensing data into process models like the DANUBIA modeling compound. In this study the focus lies exclusively on spaceborne remote sensing data.
The modifiability of areal units in remote sensing data is prevalent in several ways. First, it should be noted, that two different observations of a remote sensing instrument of the same spot may be dislocated with respect to each other due to altered orbit or orientation of the spacecraft. Thus, two successively collected measurements may be mutually shifted. In this case apparently comparable pixels capture different areas, even after georeferencing. When focusing in on a single pixel it has to be recognized that the observed region on the ground is of round or oval shape while the resulting measurement is abstracted as a square box. When comparing data of different resolution this will result in areas sampled at one resolution that are omitted at another (Figure 3.8). Yet it is common practice to assume identity of the observed areas taken as square raster elements. Thirdly, remote sensing scanners collect data by viewing across their flight track to each side of the nadir direction. Depending on the scan angle and the IFOV of a sensor, the size of the ground area sampled changes. Thus, nominal resolution of a sensor may be substantially deteriorated within a scan line while the pixel size attributed to an image remains constant. An example of this deterioration of resolution within the same scan lines is given in Figure 3.9.
Figure 3.8: Discrepancy between actual measured area of a sensor and the abstraction as pixels. The small circles and squares correspond to a high- resolution sensor; the large circle and square correspond to a low-resolution sensor. The pale shapes in the background illustrate dislocation of a measurement and pixel in data from another orbit for the low-resolution sensor.
Figure 3.9: Effect of viewing angle on resolution in MODIS imagery; left: close to nadir viewing image; right: off nadir view of the same area at a scan angle of ~40°
In the context of this study data of two sensors is sought. High-resolution data at 30m ground resolution is used from the Landsat Thematic Mapper TM sensor. Low-resolution data was collected by the Moderate Resolution Imaging Spectrometer MODIS. This instrument collects data for land surface applications at 250m and 500m sampling intervals depending on wavelength. Furthermore, the 1km domain is covered by MODIS. Data are distributed aggregated to that pixel spacing and a host of ready to use products derived at that resolution are available (e.g. Leaf Area Index). The issue of the sensors and imageries detailed characteristics is expanded on in chapter 4.
In order to use these remote sensing data and derived parameters as input for DANUBIA or comparable process models, both spatial as well as temporal resolution have to be reflected. At first glance using high-resolution data of 30m resolution seems enticing. Operation at subscale precision would be at hand for parameter retrieval as well as process modeling. However, temporal availability of these high-resolution data is low and so is spatial coverage. Landsat TM collects data of the same area at repetitive intervals of 16 days. Taking into account frequent cloud cover in mid latitudes, not more than 5-10 cloud free observations per annum can be expected. Additionally, the extent of a mesoscale catchment like the upper Danube requires eight TM images to be fully covered (compare Figure 4.7). Thus, providing regular datasets from this remote sensing device is not feasible. Contrarily, MODIS data will cover the upper Danube basin entirely at least once every day. It is obvious that the attempt to provide frequent parameter retrievals for DANUBIA has to focus on moderate resolution data.
Deriving single data values at the scale level of the 1km proxel would be straightforward. Applying the various MODIS products is the simplest option. However, the scale issues of processes below the proxel scale could not be addressed in this manner. On the other hand MODIS and other environmental optical remote sensing instruments like MERIS do provide samplings at higher levels of resolution. Yet, is the MODIS 500m or 250m resolution appropriate to comply with subscale parameterizations as discussed in Section 3.3? An error would be introduced to subscale parameters even if the higher resolution 250m data were applied. Geocomplexes or stratification of land cover would be derived at precision of at least 1%. Creating subscale information from 250m MODIS resolution however, would provide a precision of only 6.25%. Hence, a substantial possible error would be built into the stratification. It would be desirable to find another way of segmenting moderate scale remote sensing measurements in order to provide more precise subscale stratification.
Mining into the sub-pixel content of remote sensing data is common in remote sensing science. It emerges from the knowledge that moderate resolution sensors will not capture homogeneous surfaces. Heterogeneity prevails in the measurement targets. The analysis of the spectral mixture in mesoscale remote sensing data has been addressed in various studies with many recent advances into the field (Ludwig et al. 2003, Braswell et al. 2003, Lobel and Asner, 2004, Liu and Wu, 2005). The aim of such approaches is to derive the land cover types or surfaces that contribute to the mixed signal received at moderate resolution sensors.