Formulation and assessment of narrow-band vegetation indices from EO-1 hyperion imagery for discriminating sugarcane disease

Abstract

The increasing commercial availability of hyperspectral image data promotes growing interests in the development of application-specific narrow-band spectral vegetation indices (SVIs). However, the selection of the optimum SVIs for a particular purpose is not straightforward, due to the wide choice of band combinations and transformations, combined with specific application purposes and conditions. Thus, the aim of this study was to develop an approach for formulating and assessing narrow-band vegetation indices, particularly those from EO-1 Hyperion imagery. The focus of SVI development was for discriminating sugarcane areas affected by 'orange rust' (Puccinia kuehnii) disease in Mackay, Queensland, Australia. After a series of pre-processing and post-atmospheric correction techniques, an empirical-statistical approach to SVI development was designed and implemented. This included the following components: a) selection of sample pixels of diseased and nondiseased areas, b) visual examination of spectral plots to identify bands of maximum spectral separability, c)generation of SVIs, d) use of multiple discriminant function analysis, and e) result interpretation and validation. From the forty existing and newly developed vegetation indices, the output discriminant function (i.e. a linear combination of three indices) attained a classification accuracy of 96.9% for the hold-out sample pixels. The statistical analyses also produced a list of function coefficients and correlation rankings that indicate the predictive potential of each SVI. The newly formulated 'Disease-Water Stress Indices' (DSWI) produced the highest correlations. The approach designed for this study provided a systematic framework in the formulation and assessment of SVIs for sugarcane disease detection.