Soybean crop area estimation through image classification normalized by the error matrix

Abstract

The objective of this work was to estimate soybean crop area by the normalization of the error matrix generated from the supervised classification of TM/Landsat‑5 images. Eight municipalities of the state of Paraná, Brazil, were evaluated using data from the 2003/2004 crop season. Classifications were carried out using the parallelepiped and maximum likelihood methods, resulting in a “soybean mask”. Kappa index values for the eight municipalities were above 0.6. Estimated soybean areas, corrected by the error matrix, were highly correlated with official estimates of the state and with estimates generated from an alternative method called “direct expansion”. Soybean crop area estimation by the normalization of the error matrix is less costly and can aid conventional methods in estimating harvests in a less subjective manner.