In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
between March and April in the following year. According to 
IAPAR (2000) the climate in this region is subtropical humid 
meso-thermic, with warm summers (Dec-Mar). 
2.2 Meteorological data 
Meteorological data between 2000 and 2007 were collected 
from the Parana Meteorological System (SIMEPAR). Mean 
air temperature (°C) and rainfall (mm) data on a daily scale 
from five meteorological stations were used in this study: i) 
Candido de Abreu (-24.63; -51.25; 645m); ii) Cerro Azul 
(-24.81; -49.25; 66m); Hi) Jaguariaiva (-24.22; -49.67; 
900m); iv) Ponta Grossa (-25.21; -50.01; 885m); v) 
Telemaco Borba (-24.33; -50.62; 768m). 
2.3 Remote Sensing information 
2.3.1. Landsat TM images 
Twelve images from Landsat 5 and 7, Thematic Mapper -TM 
sensor were used to identify soybean fields during the time 
period 2000 to 2007. Landsat scenes were provided by INPE 
{National Institute for Space Research - Brazil) covering the 
Campos Gerais region. Two phenological stages are 
important to identify soybean crops in the Campos Gerais 
region: i) November and ii) February. In November, during 
sowing and seedling emergence, there is a strong spectral 
response from the soil, while in January/February (around 
grain maturation) the main response is from the crop 
chlorophyll activity (Rizzi and Rudorff, 2005; Wagner et al., 
2007). These authors emphasize that during January to 
March, the soybean crop is at full growth and covers the soil. 
In these conditions, soybean crops are well characterized in 
the images, standing out from other land uses such as bare 
soil or vegetation having different phenology. 
2.3.2 MODIS 250m NDVI composites 
This study used multi-temporal values of the Normalized 
Difference Vegetation Index (NDVI) obtained from the 
MODIS-TERRA sensor (MOD13Q1 product version 5, 16- 
day image composite). NDVI values for different samples in 
Campos Gerais representing “pure-pixels” of soybean, i.e., 
pixels with soybean crops only (identified using Landsat 
images as described above) were extracted using the IRI Data 
Library {http://portal.iri.columbia.edu/portal/server.pt), for 
the period 2000 - 2007. 
3. METHODOLOGY 
3.1 Soybean field mapping 
The mapping of soybean areas was performed by analyzing 
the 12 Landsat images using the following steps. First, ENVI 
4.2 software was used to geo-reference the images presented 
based on RGB color composites (using channels 4, 5 and 3 
respectively). Then, a mixture model described by 
Shimabukuro and Smith, (1991) was used to represent 
sample targets of green vegetation (soybeans in this case), 
soil, and shade. This analysis was conducted using SPRING, 
a Geographical Information System developed by INPE 
which operates using Microsoft Office Access data base. The 
method is based on the constrained least squares method and 
spectral signatures collected directly from Landsat TM 
images. This procedure was performed interactively until 
verifying that the component signatures were adequate 
representations of mixture components for the analyzed area 
Fifty-pixel samples were chosen for the test, as described in 
the procedure described by Rizzi and Rudorff (2005). The 
next step consisted of using a segmentation approach of 
growing “regions”, where a “region” is a set of homogeneous 
pixels grouped according to their spectral and spatial 
properties. An unsupervised classification based on a 
‘clustering’ algorithm, named ISOSEG (Bins et al. 1996), 
was applied to the segmented image. An acceptance threshold 
was defined (using maximum Mahalanobis distance method) 
to classify digital numbers of regions based on their closeness 
to the mean of the class (in this case, soybean gray levels). It 
differs from Euclidean distance in that it takes into account 
the correlations of the data set and it doesn’t depend on the 
scale of measurements. A thematic map containing ‘soybean’ 
and ‘non-soybean’ areas resulted from this step. In order to 
correct errors resulting from the digital classification 
described above, a thorough visual classification was 
performed in all images. Google Earth images at high spatial 
resolution were used as ancillary information to check the 
soybean fields. The Kappa index was computed to check map 
accuracy. At the end of all of these steps, one soybean field 
map was generated for each one of the seven years (2000- 
2007). A methodology was applied in order to construct a 
sampling frame and to allocate regular pixels to the soybean 
fields. A direct expansion estimator and a semi-automated 
procedure as suggested by Adami et al. (2007) were used to 
generate a grid with regular pixels of 250m by 250m. Each 
pixel was assigned an identification number. This grid spatial 
resolution was adopted to match soybean pure-pixels with the 
NDVI-MODIS pixel size. 
3.2 Soybean pure-pixel coordinates collection 
From each soybean field map and for each year, the 
coordinates of several soybean pure-pixels were extracted. In 
addition, the soybean pure-pixels closest to the 5 
meteorological stations were selected for further analysis. 
This process was needed to perform the water balance from 
the meteorological data and compare it with the NDVI. Five 
pure-pixels on the soybean field maps were chosen for each 
location and for each year, totaling 175 pure-pixel samples. 
3.3 NDVI data acquisition and processing 
NDVI values were extracted for the same 175 pure-pixels 
collected in the step described above. Then, NDVI means 
were calculated for each location and for each year. 
3.4 Rainfall anomalies and Linear Regressions analyses 
In order to verify the rainfall interannual variation during 
soybean growing season, anomaly values were calculated for 
each year separately, from 2000 to 2007. In addition, NDVI 
curves and amount of rainfall (mm) were plotted together to 
verify the correspondent variations. However, to investigate 
the relationship between NDVI, rainfall and actual 
évapotranspiration for the entire study period (2000-2007), 
regression analyses for each location were made. 
3.5 Water Balance Methodology 
The water balance model applied in this study was based on 
the methodology presented by Thornthwaite and Mather 
(1955). Thornthwaite (1948) developed a method to estimate 
Potential Evapotranspiration (PE or ETo), using air 
temperature as the main parameter which was also used in 
this study. A crop coefficient (Kc) was used to transform