Multispectral image classification from axiomatic locally finite spaces-based segmentation

Abstract : To preserve geometric features, GEOBIA-based image classification first segments a scene grouping geometrically near pixels in superpixels, with the additional benefit of meaningfully decreasing the number of spatial units to be classified. This grouping is usually based on neighbourhood definitions that use (a,b)-connected pixel graphs. Such an approach aims to find object boundaries, which are one-dimensional, from pixels, which are two-dimensional. However, this is practically impossible since topological axioms needed to ensure a correct representation of connectedness relationships cannot be satisfied by (a,b)-connected pixel graphs. The purpose of this paper is to test a novel image segmentation approach from unidimensional space elements based on axiomatic locally finite spaces (ALFS) and their linked oriented matroids. The proposed approach comprises the following stages: (i) multi-scale texture analysis using Gaussian derivative filter kernels; multi-scale gradient calculation for both multispectral image and , spectral affinity analysis and watershed transform were also used in boundary texture analysis image; geometric strength boundary; (iv) affinity analysis-based spectral strength boundary; (v) global strength boundary; (vi) watershed transform definition; (vii) directional filtering response and (viii) support vector machine (SVM) classification. We tested the proposed approach classifying a small subset of GEOBIA2016 benchmark dataset over the town of Potsdam (Germany). For the test, both the ALFS-based and conventional image segments were classified using the support vector machine (SVM) algorithm and compared their results. The use of ALFS-based watershed transform using multiscale texture and multispectral gradients and affinity analysis favoured the proper definition of super pixels resulting in a classification overall accuracy improvement. On the other hand, the general accuracy value of the classification was also improved when the directional filtering response channel was used. However, the increase is small since this response is scale dependent and therefore, resulting classes are mixed depending on the combination of objects with different scales and texture patterns. It is the case of Low vegetation and Building classes which were the two classes that became confused in both conventional and ALFS representations. Results from this study show that the ALFS-based image classification allows to improve the overall accuracy. Results also suggest that the inclusion of additional, directionally filtered bands, not always improve thematic accuracy due to the scale dependence of directional filtering. The availability of an underlying ALFS provides a topologically correct affinity analysis since, in contrast to the conventional representation based only on pixels, available inter-pixel elements allows an adequate affinity assessment along a connecting line whose representation is made from unidimensional elements. While results of the experiments show that the benefits of the inclusion of an underlying space based on ALFS are not significant, the possibility to fulfil topological requirements suggest that the ALFS-based image analysis framework is worth of further development. Authors will explore options to reformulate several processing tasks based on Cartesian complexes to produce better outcomes. This includes fitting cylindrical parabolas to elliptical patches at each unidimensional space element after computing oriented gradient to counteract the phantom order effect (i.e. to avoid producing borders which do not exist).
Complete list of metadatas

http://hal.univ-reunion.fr/hal-01960105
Contributor : Réunion Univ <>
Submitted on : Wednesday, December 19, 2018 - 10:57:29 AM
Last modification on : Tuesday, April 16, 2019 - 12:48:47 PM

Identifiers

  • HAL Id : hal-01960105, version 1

Collections

Citation

Jose Valero, Ivan Lizarazo. Multispectral image classification from axiomatic locally finite spaces-based segmentation. GEOBIA 2018 - From pixels to ecosystems and global sustainability ​, Centre d'Etudes Spatiales de la BIOsphère (CESBIO); Office national d'études et de recherches aérospatiales (ONERA); Espace pour le développement (ESPACE DEV); Société T.E.T.I.S, Jun 2018, Montpellier, France. ⟨hal-01960105⟩

Share

Metrics

Record views

41