Industrial tomato yield prediction using machine learning models

Tomato is steadily the most produced and consumed vegetable in the world during the last decade. The wide availability of data produced by the worldwide industrial tomato cultivation efforts, in combination with the power of machine learning algorithms, allows the identification of patterns and hidden variable correlations that allows the creation of accurate prediction models. These models can be used to accurately predict the yield of cultivations under various conditions and environments. In the present study, a machine learning platform was developed, able to predict the yield of industrial tomato crops during their cultivation period, based on historical data (yield, the hybrid cultivated and the environmental conditions of the area). These data were extracted from 302 different fields in six regions of western Peloponnese, Greece. They include data from 2019 to 2021 with 26 different tomato hybrids being cultivated during three cultivation periods, one per year. A correlation study was conducted on the dataset confirming that most of the variables are correlated to some degree to the crop yield. In order to find the optimal algorithm for this dataset, 12 different algorithms were tested and ranked using 14 different metrics. In addition, a hyper-parameter optimization process ensured that each tested algorithm is optimized for the dataset under examination. The presented solution is using the Ridge algorithm as it was deemed the most appropriate. The model was tested on open field cultivations, predicting the yield of tomato crops within the cultivation period, using real time weather data for 11 fields during the 2022 cultivation period. The observed prediction error ranged from just 6 up to 6,674 kg/ha, which was considered as acceptable by the producers. The uniqueness of the presented study lies in the fact that the model was both trained and tested in open field cultivations using in-season environmental data. • The dataset contained 636 records from 302 fields with 26 different tomato hybrids. • The model was trained and tested in open field cultivations using only raw data. • The prediction error ranged from just 6 kg/ha. • In-season weather data used for yield prediction in 11 open-field farms. • The predictions of the model improve as monthly data is added, approaching zero deviation.