Evaluation of Machine Learning Models for Water Stress Detection Using Stem Impedance

Abstract

Food security, producing enough food for every person on the planet, is becoming a significant issue. Increasing world population and climate change are setting new challenges to food production. Water stress can cause severe damage to crops, and detecting and preventing this threat is crucial. Smart agriculture and the use of sensors directly on the field is a promising and rapidly evolving solution. Data collected by a large number of sensors must be analyzed and efficiently interpreted. In this context, machine learning is an effective solution. This article conducts a comparative analysis of several well-established machine learning models, all trained on a dataset enriched with a novel parameter for the assessment of plant health, the stem electrical impedance (modulus and phase). This feature gives promising results since it is a direct parameter of the plant itself. Moreover, the inclusion of the stem impedance parameter significantly boosted the model's performance, notably enhancing the effectiveness, particularly evident in the case of the top-performing model in this study, the random forest algorithm. When incorporating stem electrical impedance, this model achieved an impressive F1 score of 98%, markedly surpassing the 88% obtained in its absence. As a complementary analysis, a permutation feature performance analysis was conducted, highlighting the potential of stem impedance modulus as a promising feature for evaluating plant watering conditions. The removal of impedance modulus from the training model resulted in an average classification performance loss of 25% in terms of F1 score, suggesting how impedance monitoring is a promising approach for plant health management.