INCORPORATING SPECTRAL INDICES AND TEXTURAL FEATURES FOR IMPROVED CLASSIFICATION ACCURACY TOWARDS SEMI-AUTOMATED RIVER SAND MAPPING USING HIGH-RESOLUTION MULTISPECTRAL SATELLITE IMAGERY

Abstract: For the ephemeral river channels in semi-arid regions of India, after every Monsoon
season, prompt preparation of river sand distribution maps is often necessary for river sand auditing before resuming the sand mining operations. The process can be readily assisted by
classifying the satellite-based remotely sensed imagery, although often confronted by limited
accuracy levels arising due to poor distinguishing capability among the spectrally similar class categories.
This study aims to improve the classification accuracy targeting river sand deposits by systematically examining the effects of including spectral indices and textural features in the feature-space while classification. Two most common classification algorithms, viz. Maximum
Likelihood Classification (MLC) and Support Vector Machine (SVM) classification were used.
The results show that SVM performed even better when Normalized Difference Vegetation
Index (NDVI) and correlation texture feature computed at 3×3 window size were included in the
feature-space comprising original spectral bands.

Keywords: River Sand, MLC, SVM, Spectral Indices, Textural Features

Introduction
River sand is a valuable natural resource with its major utilization in the construction industry.
With the increasing pressure of rapid infrastructural development, there has been an equivalent rise in the market demand of river sand. Consequently, the limited sources of river sand are placed with an undue strain in terms of disturbed riverine environment. To ensure the sustainable use of river sand, regular audits of their reserves are often recommended by the policymakers. The preparation of river sand distribution maps is an important step in river sand audits. Regular updating of such maps requires frequent cumbersome field surveying spells
that usually encompasses the involvement of multiple stakeholders. In this scenario, often, old data
are relied upon and are carried forward year-by-year without updating due to lack of availability
of resources, and more importantly, due to the lack of knowledge of appropriate technique (Mitra
and Singh, 2015). Remote sensing data has demonstrated potential in the applications
involving the generation of planimetric maps depicting river sand deposits. For example,
Ramkumar et al. (2015) made an attempt to identify the active-channel sand bars within Kaveri
River through visual interpretation from IRS 1B LISS-III imagery of the year 2008, compared it
with Survey of India toposheet of the year 1971, and found the areal extent of these sand bars
to be increasing at a rate of 1.05 km2 per year (Ramkumar et al., 2015).
For eliminating the problem of biasedness in image interpretation, and towards introducing automation, numerous classification algorithms have been developed by the research community such as pixel-based, object-based and knowledge-based classification algorithms that use spectral, spatial or temporal information, or any combination of these, via unsupervised clustering methods or supervised learning methods (Lu and Weng, 2007). For example, Leckie et al. (2005) performed pixel-based image classification on 0.8 m spatial resolution aerial imagery dataset with eight spectral bands for mapping of stream features such as deep, moderate and shallow glasses of water along with sand, gravel, cobble, and rocky areas. Feature-space optimization is a critical step towards increasing the classification accuracy. Features such as spectral signatures, vegetation indices, textural and terrain features are some of the possible variables in any classification process (Lu and Weng, 2007). The key is to improve the class-separability by incorporating the distinguishing characteristics among the land cover classes present in the study area. Various spectral indices, when included in the
9
feature-space, have contributed towards improving the overall accuracy of classification (Zha et al., 2003). Normalized Difference Vegetation Index (NDVI), developed by Rouse et al. (1973) is one of the commonly used indices. Also, the Normalized Difference Water Index (NDWI) is specifically used to enhance the surface water features and works as a complementary index along with NDVI (Zha et al., 2003). By using NDWI, estimation of turbidity in water may be performed (McFeeters, 1996). To enhance the separation of completely bare, sparse and dense vegetation cover, Bare Soil Index (BSI) is used (Azizi et al., 2008).
Further, incorporating image texture properties for improved classification accuracy has been suggested by many studies like Haralick et al. (1973) and Huang et al. (2014). Image texture is a measure that depicts the spatial arrangement of the pixel grey levels (Carbonneau et al., 2005). Although there are several measures available to harness textural properties from
image data (Chaurasia and Garg, 2013), Grey Level Co-occurrence Matrix (GLCM) is one
of the widely utilized methods in case of satellite imagery (Haralick et al., 1973; Huang et al.,
2014; Li et al., 2011). However, it is important to consider the choice of texture feature and window size to adopt for the study (Pathak and Dikshit, 2010). The present study attempts to make an assessment on the hypothesis that for the spatial mapping of river sand, spectral indices and textural features improve the accuracy of supervised classification on the multispectral satellite imagery. Various combinations of features were prepared and each of the combinations was examined using two most prevailing supervised classification methods,
namely, Maximum Likelihood Classification (MLC) and Support Vector Machine (SVM). MLC is a parametric classification method wherein it is assumed that the land cover features are normally distributed and hence, to derive the statistics, a sufficiently large number of training samples are required for MLC (Richards and Jia, 1999). Whereas SVM, which has sharply gained popularity in remote sensing (Mountrakis et al., 2011), is a non-parametric machine learning algorithm where parameters such as mean vector and covariance matrix are not used,
and hence, any kind of assumption about the data is not required (Lu and Weng, 2007). The reader is referred to Lu and Weng (2007) for a detailed description of various remote sensing image classification methods.

Read More

By:

Virat Arora, S. S. Rao, E. Amminedu and P. Jagadeeswara Rao
National Remote Sensing Centre, Hyderabad, India
Dept. of Geo-Engineering, Andhra University College of Engineering, Visakhapatnam, India
Email: aroravirat@gmail.com