Plant stress early detection through a low-cost multispectral device: Toward safer and more sustainable agricultural practices.

While multispectral sensors offer a cost-effective and robust solution for monitoring plant responses to environmental stress, their limited spectral resolution, largely dependent on vegetation indices, can hinder accurate classification of stress severity using machine learning. This paper aims at overcoming these limitations by introducing a multispectral device for plant stress early detection that is 1) affordable for a wide range of end-users, 2) robust to environmental factors, 3) capable of automatically finding the most meaningful features that maximize the stress detection accuracy, and 4) capable of discriminating different plant stress severity. The device integrates a broadband LED and a VIS-NIR multispectral sensor to early predict plant stress through machine learning algorithms (i.e., SelectKBest, kNN, SVM, and LDA). It was trained on spectral measurements acquired from tobacco plants under salinity stress. The results demonstrated its high capability to discriminate with high accuracy different stress severity (average accuracy of 91.0 ± 3.1%).