Classification of Agricultural Fields in Satellite Images Using Two-Dimentional Hidden Markov Models

BAUMGARTNER JOSEF; GIMENEZ ROMERO, JAVIER ALEJANDRO; PUCHETA JULIAN; FLESIA ANA GEORGINA

Córdoba

Congreso; Congreso Argentino de AgroInformática; 2013

Image segmentation is a key competence for many real life applications such as precision agriculture. In this work we present an approach to classify agricultural elds in noisy satellite images. We start with the Markovian neighborhood hypothesis from where on we derive a general two-dimensional hidden Markov model (2D-HMM). To make the 2D-HMM feasible we apply the Path-Constrained Variable-State Viterbi Algorithm (PCVSVA) which allows us to approximate the optimal hid- den state map. We evaluate the PCVSVA for a Landsat image of the province of Cordoba, Argentina and a synthetic satellite image. In both cases we use Cohen's b coecient to compare the PCVSVA and the so- lution obtained by maximum likelihood (ML) to show the eectiveness of 2D-HMM of solving image segmentation tasks.