Volume XLI-B7
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLI-B7, 205-211, 2016
https://doi.org/10.5194/isprs-archives-XLI-B7-205-2016
© Author(s) 2016. This work is distributed under
the Creative Commons Attribution 3.0 License.
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLI-B7, 205-211, 2016
https://doi.org/10.5194/isprs-archives-XLI-B7-205-2016
© Author(s) 2016. This work is distributed under
the Creative Commons Attribution 3.0 License.

  21 Jun 2016

21 Jun 2016

A METHOD TO ESTIMATE TEMPORAL INTERACTION IN A CONDITIONAL RANDOM FIELD BASED APPROACH FOR CROP RECOGNITION

P. M. A. Diaz1, R. Q. Feitosa1,3, I. D. Sanches2, and G. A. O. P. Costa3 P. M. A. Diaz et al.
  • 1Pontifical Catholic University of Rio de Janeiro, Brazil
  • 2National Institute for Space Research, Brazil
  • 3Rio de Janeiro State University, Brazil

Keywords: Conditional Random Fields, Crop Recognition, Multitemporal Image Analysis

Abstract. This paper presents a method to estimate the temporal interaction in a Conditional Random Field (CRF) based approach for crop recognition from multitemporal remote sensing image sequences. This approach models the phenology of different crop types as a CRF. Interaction potentials are assumed to depend only on the class labels of an image site at two consecutive epochs. In the proposed method, the estimation of temporal interaction parameters is considered as an optimization problem, whose goal is to find the transition matrix that maximizes the CRF performance, upon a set of labelled data. The objective functions underlying the optimization procedure can be formulated in terms of different accuracy metrics, such as overall and average class accuracy per crop or phenological stages. To validate the proposed approach, experiments were carried out upon a dataset consisting of 12 co-registered LANDSAT images of a region in southeast of Brazil. Pattern Search was used as the optimization algorithm. The experimental results demonstrated that the proposed method was able to substantially outperform estimates related to joint or conditional class transition probabilities, which rely on training samples.