SPECTRAL AND TEXTURAL ANALYSIS FOR IMPROVING
CROP CLASSIFICATION ACCURACY
Francisco Redondo
Maria Cristina Serafini
Miriam Antes
Adrian Benitez
Universidad Nacional de Lujan
Dept. Ciencias Basicas
Ruta 5 y Ruta 7
Lujan - Provincia de Buenos Aires 6700
Argentina
ISPRS Commission VII / Working Group 3
ABSTRACT
The addition of low pass filtered images and textural features significantly improved class separability
when compared to a spectral only classification. Low pass filtering removes high frequency image
infornation and decreases the spread within each class signature. In consequence class overlapping
decreases. On the other hand, textural information could be used only if the representative spatial
variations of each cover are equal or higher than the spatial resolution of the sensor. Using TM and SPOT
data and the procedures above described, an experience in a representative pilot area of the main
agricultural region in Argentina was carried out. Results show an increase of more than 10 percent in the
overall classification accuracy within the training sites. Another interesting find is that the accuracy was
not decrease outside the training areas when low pass filtered images were used in the analysis. That
improvement in the classifier consistency is very important when large areas are being studied.
046