Developing supervised machine learning algorithms to classify lettuce foliar tissue samples into interpretation zones for 11 plant essential nutrients

Abstract

Greenhouse crop nutrient management recommendations based on foliar tissue testing rely heavily on human interpretation, which can result in recommendation variations and errors. Critical nutrient ranges vary for each species, and the potential for error in interpretation increases due to this complexity. Machine learning can be utilized to develop algorithms to accurately classify new information using models developed on known data from a training dataset. This study examines four different machine learning algorithms (J48, random forest [RF], sequential minimal optimization [SMO], and multilayer perceptron [MLP]) by two different cross‐validation strategies (10‐fold and 66% split) to determine if machine learning can be utilized to accurately classify foliar tissue samples within corresponding nutrient ranges. Lettuce (Lactuca sativa L.) foliar tissue samples (n = 1950) from a variety of controlled experiments and diagnostic samples from state and private labs were compiled and assigned to one of five nutrient ranges of deficient, low, sufficient, high, or excessive for each of 11 plant essential nutrients of interest based on Gamma or Weibull distributions. Individual machine learning algorithms were developed for each nutrient. For all examined essential nutrients, J48 or RF yielded the >98% greatest percentage correct classification when compared to MLP or SMO. This study establishes the novel use of machine learning for lettuce foliar nutrient analysis results interpretation with a higher accuracy rate than by traditional statistical methods.