A first conclusion, resulting from measuring the spectral response at leaves level, showed that stress was a property of plants since different leaves with different levels of damage yielded the same spectral response. Then, to assess the applicability of unsupervised classification of plants as healthy, biotic-stressed or abiotic-stressed, feature extraction and selection from leaves spectral signatures, combined with a Supported Vector Machine classifier was designed. Optimization of SVM parameters using grid search with cross-validation, along with classification evaluation by ten-folds cross-validation showed a correct classification rate of 95%, consistently on both seasons. Controlled experiments using cages with different numbers of caterpillars--including caterpillar-free plants--were also conducted to evaluate consistency in trends of the spectral response as well as the extracted features.
ACCESS THE FULL ARTICLE
Matías Tailanián, Enrique Castiglioni, Pablo Musé, Germán Fernández Flores, Gabriel Lema, Pedro Mastrángelo, Mónica Almansa, Ignacio Fernández Liñares, Germán Fernández Liñares, "Early pest detection in soy plantations from hyperspectral measurements: a case study for caterpillar detection," Proc. SPIE 9637, Remote Sensing for Agriculture, Ecosystems, and Hydrology XVII, 96372I (14 October 2015);