Estimation of thrips (Fulmekiola serrata Kobus) density in sugarcane using leaf-level hyperspectral data

Abstract

Sugarcane thrips, Fulmekiola serrata (Kobus) (Thysanoptera: Thripidae), is a recent pest in the South African sugarcane industry. First identified in South Africa in 2004 (Way et al. 2006), it is now widespread throughout the sugarcane-growing regions of the country (Keeping et al. 2008). Thrips are small insects (2−3 mm long) that feed on the spindle leaves of sugarcane. Damage caused by this thrips includes leaf necrosis due to puncturing of the leaf surface (Way et al. 2006). Younger crops tend to be most vulnerable and appear to depend on numbers of thrips present (Keeping et al. 2008). Monitoring thrips involving leaf sampling and laboratory analysis before treatment is expensive and labour intensive. Therefore, complementary methods that can provide up-to-date information are needed. Remote sensing offers timely data that has potential for sugarcane thrips monitoring as demonstrated by Mirik et al. (2007) for a similar pest of winter wheat. The use of hyperspectral data for this purpose seems particularly promising. Such data are characterised by light reflectance from many (typically several hundred), narrow, contiguous wavebands across the spectrum. Hyperspectral data are able to detect nuanced differences that could be related to different types of stress, pest infestations or disease incidences (Lillesand and Kiefer (2001). However, in order to analyse such large sets of spectral data for model development, one would need to collect many sample data to avoid overfitting (high variable-to-sample ratio problem). The collection of many such data is often impossible due to logistical and other constraints. Therefore, researchers seek techniques and methods that could be used to reduce the redundancy and colinearity in the hyperspectral data without losing information that is relevant to the features of interest. Random forest, a machine learning algorithm developed by Breiman (2001), is a relatively new method that has been used for such a purpose (Chan and Paelinckx 2008, Ismail 2009). The random forest regression method uses several user-defined parameters and random selection of input variables to predict a feature of interest (Breiman 2001, Maindonald and Braun 2006). The method provides information about the importance of the variables on the performance of the predictive model (Breiman 2001, Archer and Kimes 2008). This can be very useful in the selection of spectral variables when hyperspectral data are analysed. One drawback of the random forest algorithm in selecting variables from the spectroscopic data is that the selected relevant wavebands could still be autocorrelated (Strobl et al. 2008), especially with those of very high spectral resolutions of handheld hyperspectral sensors. Partial least squares (PLS) regression (Wood et al. 1996) overcomes the colinearity problem (Huang et al. 2004), but one issue with PLS regression in hyperspectral data analysis is the identification of the most influential spectral region(s) during the models development (Huang et al. 2004). Martin et al. (2008) recommended that the PLS regression coefficients be normalised by the average spectral reflectance at all input wavebands. Spectral regions that show high values of normalised coefficients indicate the influence of such regions on the calibrated PLS regression models