, < 
In: Wagner W„ Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
570 
the area suitable for the cultivation of vines, tobacco (main cul 
tivation in the region), fruits and grains. Slopes are deforested 
and eroded; with areas prone to landslides. Most of the hills are 
covered by low-productivity grassland used for grazing. There 
are alluvial and deluvial-meadow soils along the major rivers in 
the region, where vegetables and hemp can be grown, due to the 
larger quantity of moisture, they receive from the soil layers. 
4 REMOTE SENSING IMAGERY 
A constellation of 5 multispectral satellite sensors were launched 
by RapidEye in August 2008 with a primary focus on agricultural 
applications. These satellites have a lifespan of seven years; a 
ground sampling distance of 6.5m resampled to 5m; and a daily 
overpass. A new feature in RapidEye sensor is the Red Edge band 
(690-730nm), which could allow better estimation of the ground 
cover and chlorophyll content of the vegetation (Haboudane et al., 
2002, Vinal and Gitelson, 2005). All 5 satellites have the same 
calibration coefficients. The radiometric scale factor converting 
the image DN values into reflectance is 0.01. 
Imagery was obtained from RapidEye AG at standard processing 
level 3A 2 (orthorectified) for the dates indicated in Table 1. Pre 
processing of imagery was carried out in ERDAS Imagine and 
ESRI ArcGIS software. This entailed further geo-referencing of 
the satellite imagery to the national orthoimagery provided by the 
Bulgarian government, thus ensuring data consistency between 
the RapidEye imagery and the LPIS datasets. Nearest neigh 
bour approach was used for the resampling. In addition to the 
RapidEye imagery, VHR data from IKONOS has been acquired 
in the frame of the annual CwRS campaign and was also provided 
for the study. The availability of this imagery was an important 
source of ground truth. An orthorectification of this VHR data 
was carried out using the reference national orthophoto, addi 
tional ground control points and the SRTM DEM provided freely. 
Acquisition dates in 2009 
Zone 
April 
May 
June 
July 
September 
12.04 
20.05 
10.06 
15.07 
10.09 
KARD 
23.07 
16.09 
24.07 
Table 1 : Acquisition dates of all RapidEye images over KARD 
5 METHODOLOGY 
The proposed methodology is based on the key elements derived 
from the GAC definition in Section 2. From the adopted GAC 
definition, we can conclude that, a land could be considered in 
GAC, if at least the following two criteria are met: i) vegetation 
is growing or can be grown on that land; ii) the land is accessible 
for agriculture activities (cropping, grazing, etc.). Both criteria 
can be evaluated by monitoring the development of the vegeta 
tion during the year (phenological cycle), together with the anal 
ysis of the texture properties of the land cover and the relevant 
spatial context. Thus, the methodological approach was based on 
a multi-temporal analysis of RapidEye time-series, using object 
oriented classification techniques in order to detect and qualify 
the land cover features in respect to their potential to represent 
agriculture area in GAC. Considering that the proposed definition 
of GAC is quite broad, it was agreed that the first estimation of the 
land potentially useful for agriculture (and being in GAC) can be 
done through detection and quantification of the non-GAC (and 
potential non-GAC) features. From land cover (physiognomic- 
structural) point of view, land which is not in GAC is constantly 
2 http ://www. rapideye.de/home/products/standard-image- 
products/standard-image-products.html 
bare or non-vegetated during the (cultivation) year (for example 
sealed surfaces; natural bare areas) and contains features prevent 
ing the agricultural activity even though it is vegetated (for exam 
ple closed forest, woodland, wetland, etc.) 
An overview of the proposed methodology used for decision 
making and analysis can be seen in Figure 2. The selection, ac 
quisition and pre-processing of imagery was important to provide 
a solid foundation for future analysis. The acquisition windows 
were carefully defined on the base of crop calendars, provided 
by ReSAC. Imagery from April, May, June, July and Septem 
ber were acquired over the test zones to reflect the phenological 
cycles of the vegetation (see Table 1). 
Figure 2: Proposed methodology 
Data analysis: Capture and qualification of the permanently non- 
vegetated areas and the areas not accessible for agriculture were 
the primary targets. Most of the non-vegetated area with artifi 
cial anthropogenic origin, as urban structures, roads, hardpans, 
etc. can be efficiently extracted from a single RapidEye image, 
if acquired in the correct period of the year. The same is also 
valid for the naturally vegetated areas, not suitable for agricul 
ture activities (such as forested areas, wetlands), or water bodies. 
However, for most of the natural bare areas, unsuitable for agri 
culture (such as as eroded surfaces, degraded soils), the analysis 
has to be based on several time series, in order to filter out tem 
porary bare areas, for example, harvested agricultural fields. An 
other important land cover, requiring multi-temporal approach, is 
low-productivity grassland, quite common in the area of KARD. 
This type of mountain grassland, used for grazing, appears vege 
tated (on the satellite imagery) only during particular periods of 
the year (early spring); usually its spectral signature is similar to 
bare surface, especially during the summer when it becomes dry. 
The only option to efficiently capture those areas is through the 
use of multi-temporal imagery covering the entire active agricul 
ture period so that various aspects of the vegetation growth and 
climatic conditions can be considered. 
It was assumed that permanent bare areas should have low NDVI 
values in all time series. For that purpose Red Edge Normal 
ized Difference Vegetation Index (NDVI) (equation 1) (Wu et al., 
2009) was calculated for all images. 
NDVI Red Edge = 
NIR — Red Edge 
NIR + Red Edge 
(1) 
The choice of the Red Edge channel, instead of the Red channel 
for the NDVI calculation, was mainly driven by lesser saturation 
of the Red Edge NDVI comparing to the traditional NDVI over 
highly vegetated (forested) regions, as reported in the literature 
(Haboudane et al., 2002, Vinal and Gitelson, 2005). Figure 4.b