A method for real-time classification of insect vectors of mosaic and brown streak disease in cassava plants for future implementation within a low-cost, handheld, in-field multispectral imaging sensor

Abstract Background The paper introduces a multispectral imaging system and data-processing approach for the identification and discrimination of morphologically indistinguishable cryptic species of the destructive crop pest, the whitefly Bemisia tabaci. This investigation and the corresponding system design, was undertaken in two phases under controlled laboratory conditions. The first exploited a prototype benchtop variant of the proposed sensor system to analyse four cryptic species of whitefly reared under similar conditions. The second phase, of the methodology development, employed a commercial high-precision laboratory hyperspectral imager to recover reference data from five cryptic species of whitefly, immobilized through flash freezing, and taken from across four feeding environments. Results The initial results, for the single feeding environment, showed that a correct species classification could be achieved in 85–95% of cases, utilising linear Partial Least Squares approaches. The robustness of the classification approach was then extended both in terms of the automated spatial extraction of the most pertinent insect body parts, to assist with the spectral classification model, as well as the incorporation of a non-linear Support Vector Classifier to maintain the overall classification accuracy at 88–98%, irrespective of the feeding and crop environment. Conclusion This study demonstrates that through an integration of both the spatial data, associated with the multispectral images being used to separate different regions of the insect, and subsequent spectral analysis of those sub-regions, that B. tabaci viral vectors can be differentiated from other cryptic species, that appear morphologically indistinguishable to a human observer, with an accuracy of up to 98%. The implications for the engineering design for an in-field, handheld, sensor system is discussed with respect to the learning gained from this initial stage of the methodology development.