Comparison of selected textural features as global content-based descriptors of VHR satellite image

Abstract

The paper presents the results of the study of usefulness of selected textural features for describing the content of image chips (1024 by 1024 pixels) cut from WorldView-2 panchromatic images. Several texture analysis techniques were applied to calculate 290 different global image descriptors. These features together with 4 histogram-based characteristics formed the set of classification features. Decision trees were used to classify the images into broad land cover categories. Performance of each group of textural characteristics was assessed and compared to the results of classification based on all calculated characteristics. The highest classification accuracy (99.6%) were achieved based on all computed attributes. Using only three features selected based on classification algorithm feature reduction capabilities, the accuracy of 99.2% was obtained. When the performance of particular texture analysis techniques is compared, the highest classification accuracy (98.8%) was achieved based on multifractal parameters used together with histogram-based features. The autoregressive model parameters gave the best classification accuracy (97.9%) when solely textural features of particular groups were used.

What Can Multifractal Analysis Tell Us about Hyperspectral Imagery?

“…Coefficient values were much lower for the three other landscape types. In general, the comparison performed here of statistical and multifractal description of hyperspectral data states the confirmation our previous analysis with panchromatic bands ( [26,35,47]) and indicates that ∆ provides complementary information to moments and may be used as an additional classification feature. This is in agreement with the fact that multifractal analysis considers both positive and negative higher-order moments (see Section 3), while in the case of statistical description, only positive or particular moments are used.…”

“…A number of experiments have shown, however, that a fractal formalism may not be sufficient for satellite imagery description, and that multifractals (an extension of fractal theory) should be used instead (e.g., [24][25][26]). The multifractal approach is based on the assumption that it is necessary to use a number of non-trivially connected fractals, each with a different dimension of self-similarity, to describe data complexity.…”

“…A review of existing implementations of a multifractal formalism in the analysis of remote sensing images indicates that they are mostly related to the segmentation and classification of radar imagery. Other applications include, for example, the detection of changes ( [16,31]), the edge-preserving smoothing of high-resolution satellite images [32], super-resolution image analysis [33], compression of remote sensing imagery [34] and usage as textural features for content-based image retrieval [35]. There are also combinations with methods such as granulometric analysis [36] or principal component analysis [37] (see Table 2).…”

“…There are also combinations with methods such as granulometric analysis [36] or principal component analysis [37] (see Table 2). Our previous research comprehensively analysed the multifractal character of panchromatic very high resolution (VHR) satellite images acquired by the following satellites: EROS [26], WorldView-2 [25], GeoEye-1 and Pléiades 1A [45,46], using the box-counting based moment method. In particular, we proposed a quantitative parameter, named the degree of multifractality (∆), and demonstrated its usefulness in attempts to distinguish the basic forms of land cover (water, urban areas, forests and agriculture).…”

“…In our previous work, we calculated the degree of multifractality for several subset sizes: 1024 × 1024 in [25], 512 × 512 in [22,26,47], 256 × 256 in [16,46] and 64 × 64 and 32 × 32 in [45]. These earlier studies indicated that the bigger the size, the more reliable the multifractal scaling and ∆ estimations.…”