Fuzzy Classification of Aster Data for Forestry Mapping

Fuzzy Classification of Aster Data for Forestry Mapping

SHARE


Syed Sohel Ali
School of Geography
Population & Environmental Management
Flinders University, GPO Box 2100, Adelaide, SA 5001
Tel: + 61 8 8182 4000, Fax: +61 8 8285 6710
[email protected]

Paul Dare
School of Geography
Population & Environmental Management
Flinders University, GPO Box 2100, Adelaide, SA 5001
Tel: + 61 8 8182 4000, Fax: +61 8 8285 6710
[email protected]

Abstract
New sensors and new technologies in remote sensing have provided a useful means of data acquisition for mapping forest environments. In this paper, a new dataset from the Advance Spaceborne Thermal Emission and Reflection Radiometer (ASTER) was used with fuzzy supervised classification. A case study of Albany, Western Australia demonstrates that ASTER data with fuzzy supervised classification is suitable for forest mapping. In fuzzy supervised classification, ASTER visible near infrared bands (3 bands) provided the highest spectral separability among forest land cover classes and accordingly, these bands were used to represent a fuzzy set. The algorithm consisted of two parts: the estimation of fuzzy membership from training data and the development of fuzzy signatures. The partial membership of pixels were determined using training data and their relative frequency distribution in the histograms. This allowed the identification of the mixed pixels from component cover classes. Accuracy assessment shows that the proposed techniques are reliable for forest mapping with high classification accuracy (overall 90.96%) and with little confusion (average omission error 8.7%).

1. Introduction
New sensors and advance techniques for image processing in remote sensing are providing capabilities for mapping and monitoring forest ecology as never before. New sensors, with higher spatial resolution and finer spectral and radiometric resolution, make data acquisition easier and even cheaper than before (Davis and Simonett,1991; Sabins, 1987). The introduction of advanced image classification techniques has also enhanced the accuracy of land cover classes. Various classification approaches have been successfully applied to the analysis of remote sensing images for the past two decades. The computer-assisted classification techniques are conventionally grouped into two broad categories: supervised and unsupervised approaches (Lillesand and Kiefer, 1994). Traditionally both approaches generate one-pixel-to-one-class mapping. However, such an approach is becoming more difficult as the areas of interest contain a mixture of land cover classes. The complex land surface often causes mixed pixels in the remote sensing image, if the image pixel size is not fine enough to catch the spectral response from only a single land class. For example, a mixed pixel may contain the spectral responses from both grass and underlying soils. There is not a well-specified criterion for distinguishing between these two cover-types. The impreciseness also results from natural variations or arise through original measurements, as well as data processing (Wang,1990). In conventional image processing, a pixel can be assigned a signal attribute with respect to a given cover class. Imprecise land cover classes are now classified by introducing concepts and tools of fuzzy set theory in remote sensing.

In this paper, a new technique is proposed for ASTER images and fuzzy supervised classification. A statistical histogram method is used to derive membership grades of forest cover classes from training data. These grades were then used in a fuzzy partition matrix to create fuzzy signatures. Last of all, using these signatures the fuzzy classification was carried out with ASTER three visible near infrared bands.

2. ASTER
Images from the Advance Spaceborne Thermal Emission and Reflection Radiometer (ASTER) sensor are now available with higher resolution and more variety of spectral and radiometric resolution than contemporary multi-spectral sensors. ASTER is a research facility instrument provided by the Ministry of International Trade and Industry (MITI), Tokyo, Japan launched on NASA’s Earth Observation System (EOS) satellite called Terra (previously AM1) platform in December 1999. The primary science objective of the ASTER mission is to improve understanding of the local and regional-scale processes occurring on or near the Earth’s surface and lower atmosphere, including surface-atmosphere interactions (Yamaguchi et al., 1998).

The specification of the 14 ASTER spectral bands is shown in Table 1. Each scene covers 60 x 60km and about 600 scenes are captured daily. The recorded data exceed the specified signal to noise ratio. This instrument has three separate optical subsystems: the visible and near-infrared (VNIR) radiometer, shortwave-infrared (SWIR) radiometer, and thermal infrared (TIR) radiometer.

Table 1: ASTER baseline performance

ASTER images of path 602 and row 67 & 68 were collected from Satellite Remote Sensing Service (SRSS), Department of Land Administration (DOLA), Western Australia. These images were acquired on January 16, 2001 and were a Level-1B data product. Before classifying the data, the scenes were geo-referenced using existing 1995 Landsat TM5 image. The details of the geo-referencing and radiometric calibration is described in Syed and Corner (2003).

3. Fuzzy supervised classification
The concept of fuzzy sets (Zadeh, 1965) is not new. It was developed in late1960s as a theory to deal with imprecise information. Fuzzy logic using graded or qualified statements rather than ones that are strictly true or false. In conventional image processing techniques, a land cover phenomena is considered to exist in discrete classes. Each pixel is allocated simply to the class with which it has the greatest level of similarity. Because only one class is associated with each pixel in the output and no indication of a relative strength of class membership is provided, full membership of the allocated class is implied. However, full membership of the allocated class is often not the case. For instance, mixed pixels may, dependent on the characteristics of the classes on the ground and the spatial resolution of the imagery, be common. Furthermore, many geographical phenomenon do not rest in discrete classes but instead lie along continua, and the classes inter-grade (Leung, 1987; Wang, 1990).