The role of machine learning on Arabica coffee crop yield based on remote sensing and mineral nutrition monitoring

Abstract

Coffee yield variation in the field can be learned to obtain useful information for coffee management. Machine learning algorithms were evaluated to determine Arabica coffee yield. The Classification and Regression Tree (CART) rpart1SE algorithm used for classification and regression provided crucial nutrient thresholds to obtain high yield and minimise yield variability, by increasing vigour in well-nourished plants, with fertiliser management in the field. The use of the random forest model enabled to detect the most important variables for predicting coffee yield, as well as to identify how the nutritional status of the plants, such as Mg, Fe and Ca contents can be balanced to maximise yield. Variables related to the coffee nutritional status were more important than remote sensing variables for estimating coffee yield in the field. Despite the better accuracy of the random forest model (rf) to predict coffee yield when compared to the rpart1SE model, the particularity of each machine learning algorithm modelling was used in terms of the benefits of the results of each methodology synergistically in favour of wisely defining the best strategy and tactics for the crop management. In general, Mg leaf content was the most important variable for yield class prediction in both the 2005 and 2006 harvests by the rf model. The CART algorithm defined Mg leaf content threshold <3.615 g kg−1 in 6/15/2005 for yield classification in the first node and this threshold obtained is consistent with the available literature associated with high coffee yields between 3.6 and 4.0 g kg−1.