Semi-automatic choice of scale-dependent features for satellite SAR image classification

https://doi.org/10.1016/j.patrec.2005.08.005Get rights and content

Abstract

In this work we compare two different approaches to the use of multiple scales in the classification process of satellite SAR images. These are (I) the multi-scale co-occurrence texture analysis and (II) the semivariogram approach. Moreover, we propose a scheme for optimizing the co-occurrence window size and the semivariogram lag distances in terms of classification accuracy performance. To improve the results even further, we introduce a methodology to compute the co-occurrence features with a window consistent with the local scale, provided by the semivariogram analysis.

Examples of satellite SAR image segmentation for urban area characterization are shown to validate the procedure.

Introduction

In remote sensing data analysis the spatial neighborhood of a pixel may contain even more information than the pixel itself; even more so for SAR data, where the large variability due to the speckle noise makes the single pixel value unreliable. This implies a need to define the size of such neighborhood. Moreover, the resolution factor is also critical, since the ground size of a pixel implies in many environments that a mixed spectral response is recorded. Both the spatial neighborhood size and the mixed pixels are matters of scale, and therefore the so-called scale analysis is mandatory for most applications, starting from the very simple step of classification.

The concept of scale in remote sensing images dates back to the pioneering work of Quattrochi and Goodchild, 1997, Ramstein and Raffy, 1989, Serra, 1982, Woodcock et al., 1988. In particular, Cao and Lam (1997) introduce four different definitions of scale, depending on the cartographic, geographic, operational and measurement point of view. In this paper we refer to the global scale (GS), defined as the mean size of the objects to be recognized in a remotely sensed scene. Differently, the local scale (LS) refers to the mean textural scale in a local neighborhood of a pixel, whose size is limited by the above mentioned global scale.

Anyway, scale is related to spatial relationships, and many methods have been proposed in technical literature to provide a measure of such relationships between neighboring pixels, using the co-occurrence matrix, wavelets, the Gabor filters, semivariograms and Markov random fields. In this work we focus only on two of them, namely the co-occurrence texture analysis (Haralick et al., 1973) and the semivariogram analysis (Matheron, 1963). Co-occurrence textural features gauge the statistical properties of patterns that jointly occur in the neighborhood of each pixel. Similarly, the semivariogram provides information about textures by computing the mean squared differences between grey levels of pixels for a wide range of lag distances.

Interestingly enough, these two methods have been compared in some papers for remote sensing data classification. For example, in (Carr and Pellon de Miranda, 1998) this comparison provides interesting clues on the choice of the best approach. The authors claim to have obtained better results with radar (SIR-B) images using semivariograms and suggest a possible reason in the fact that texture in the microwave domain may not obey a Markov law. Our opinion, better explained in next sections, is that co-occurrence computation needs a careful tuning with respect to the LS of the texture in the radar image. This is also stressed by the results for SAR mapping of urban areas in (Dekker, 2001). In particular, in this work the joint use of single GS co-occurrence features and the semivariogram function leads to good classification results. This suggests that a more detailed comparison of co-occurrence and semivariogram approaches may be very interesting, and it would be particularly useful to combine, in a context-dependent way, the information from both methods.

To this aim, we should note that semivariograms have been used also for the definition of the GS in remote sensing data. Towhsend and Justice (1988) show that the semivariogram plot exhibits a maximum for the lag distance that corresponds to the mean repetition distance between basic texture elements. The method has been compared by Chen and Blong (2003) with a faster, but rougher, approach based on wavelet transform, and this can be considered as a further validation of the semivariogram GS detection capabilities.

In this work we first perform a detailed comparison between co-occurrence textures with multiple window sizes and semivariograms in a range of lag distances for satellite SAR classification. Subsequently, we introduce a new methodology, based on semivariograms to steer the local co-occurrence texture measure computation based on the LS; in particular, the local size of the window is bound to the LS of the scene. The rationale for using this strategy is to save CPU time and memory; the use of image-wide, multiple window size features, as more precisely proposed in (Dell’Acqua and Gamba, 2003), is very expensive in terms of computing resources. The proposed adaptive size scheme should be more provident.

Section snippets

Multi-scale analysis in co-occurrence and semivariogram classification of remotely sensed data

One of the most significant problems in dealing with image segmentation, and notably with remotely sensed data segmentation, is that more local scales co-exist in the same image, and often change through the scene. These values are important for a contextual classification of the image, but one single global scale falls short of capturing the information we need for a correct interpretation of the whole image. This is the reason why many authors (Pesaresi and Benediktsson, 2001, Zhang et al.,

Experimental results

The algorithms presented in the previous section have been applied to a slant range complex satellite SAR images of the area around the town of Pavia, Northern Italy. The first one is a C-band, VV-polarization image recorded by the ERS-1 satellite on the 13th August 1992, while the second one is an ASAR Alternate Polarization (HH/VV) image recorded by the ENVISAT-1 satellite on 29th August 2003. Both images depict an area of nearly 64 km2. Using a manual procedure, the images were co-registered

Conclusions

This paper addresses the use of spatial scale information as an aid to classify remotely sensed SAR data. A novel approach combining two widely used techniques, namely co-occurrence matrix and semivariogram analysis, was proposed and tested for mapping urban density classes in satellite SAR data. The results show that the joint use of co-occurrence textural features and semivariogram analysis to optimize co-occurrence window size can be, in terms of result accuracy, nearly as effective as a

References (19)

  • C.E. Woodcock et al.

    The use of variogram in remote sensing and simulated image, II: Real digital images

    Rem. Sens. Environ.

    (1988)
  • D.G. Barber et al.

    SAR sea ice discrimination using texture statistics: A multivariate approach

    Photogramm. Eng. Rem. Sens.

    (1991)
  • C. Cao et al.

    Understanding the scale and resolution effects in remote sensing and GIS

  • G.A. Carpenter et al.

    ART neural networks for remote sensing: Vegetation classification from Landsat TM and terrain data

    IEEE Trans. Geosci. Rem. Sens.

    (1997)
  • J.R. Carr et al.

    The semivariogram in comparison to the co-occurrence matrix for classification of image texture

    IEEE Trans. Geosci. Rem. Sens.

    (1998)
  • K. Chen et al.

    Identifying the characteristic scale of scene variation in fine spatial resolution imagery with wavelet transform-based sub-image statistics

    Int. J. Rem. Sens.

    (2003)
  • Dekker, R.J., 2001. Texture analysis of urban areas in ERS SAR imagery for map updating. In: Proc. 1st IEEE/ISPRS Joint...
  • Dell’Acqua, F., Gamba, P., 2003. Discriminating urban environments using multi-scale texture and multiple SAR images....
  • P. Gamba et al.

    Improved multiband urban classification using a neuro-fuzzy classifier

    Int. J. Rem. Sens.

    (2003)
There are more references available in the full text version of this article.

Cited by (45)

  • Analyzing fine-scale wetland composition using high resolution imagery and texture features

    2013, International Journal of Applied Earth Observation and Geoinformation
    Citation Excerpt :

    That said, determining the appropriate window size for certain scales is difficult (Dell’Acqua et al., 2006). Semivariograms have been extensively used in remote sensing studies to calculate optimal neighborhood size (Franklin et al., 1996; Dell’Acqua et al., 2006; Johansen et al., 2007; Balaguer et al., 2010). For example, Franklin et al. (1996) used semivariogram range to estimate and select the appropriate and optimal window size for different vegetation classes in an image.

  • Multi-sensor data fusion for estimating forest species composition and abundance in northern Minnesota

    2011, Remote Sensing of Environment
    Citation Excerpt :

    Hence, the use of PALSAR data in combination with C-band SAR and optical remote sensing data is recommended for distinguishing spruce budworm host tree species in northeastern Minnesota. Extraction of SAR amplitude statistics (C- and L-band), according to pixel-wise, optical, Euclidean neighborhoods (see Wolter et al., 2009), provides a context within which otherwise noisy SAR data (Dell'Acqua et al., 2006) can be used in conjunction with optical data (via PLS regression) to distinguish different species and facilitate estimation of their relative abundance. While results (Fig. 7) seem to suggest leaf-off Radarsat-1 data (three variables selected) are more useful than leaf-on data in this regard (two variables selected), ratios of summer to winter (leaf-on/off) Radarsat-1 data were clearly important for reconciling relative BA for several optically similar conifer species, especially spruce budworm host tree species (Fig. 7).

  • Definition of a comprehensive set of texture semivariogram features and their evaluation for object-oriented image classification

    2010, Computers and Geosciences
    Citation Excerpt :

    Thus, the range of the semivariogram has been proven to be directly related to the size of objects or patterns in an image (Curran, 1988; Woodcock et al., 1988a, 1988b). The range of the semivariogram has been used to predict optimal dimensions, such as the pixel size for remote sensing studies (Rahman et al., 2003), the dimensions of windows or neighbourhoods to be used in texture analysis (Franklin et al., 1996; Dell’Acqua et al., 2006; Johansen et al., 2007), or the local filter size for tree location from high spatial resolution imagery (Wulder et al., 2000). In forest structure analysis and forest management applications using remote sensing data, for instance, the range and the sill of the semivariogram have been used to estimate the forest structure (St-Onge and Cavayas, 1997), and later for modelling individual tree crown and forest canopy structure and health (Lévesque and King, 2003).

  • Rapid mapping of high resolution SAR scenes

    2009, ISPRS Journal of Photogrammetry and Remote Sensing
    Citation Excerpt :

    The scale of the textures and its orientation is however important to avoid erroneous extraction. The first issue is discussed in Gamba et al. (2006) for coarse resolution SAR data, while the second one has been solved for optical data using multiple directions in Pesaresi et al. (2007). A different approach, proposed in this work, is to look at the range of values in a window around each pixel.

View all citing articles on Scopus
View full text