Prediction of plant-level tomato biomass and yield using machine learning with unmanned aerial vehicle imagery

Abstract

Abstract Background The objective of this study is twofold. First, ascertain the important variables that predict tomato yields from plant height (PH) and vegetation index (VI) maps. The maps were derived from images taken by unmanned aerial vehicles (UAVs). Second, examine the accuracy of predictions of tomato fresh shoot masses (SM), fruit weights (FW), and the number of fruits (FN) from multiple machine learning algorithms using selected variable sets. To realize our objective, ultra-high-resolution RGB and multispectral images were collected by a UAV on ten days in 2020’s tomato growing season. From these images, 756 total variables, including first- (e.g., average, standard deviation, skewness, range, and maximum) and second-order (e.g., gray-level co-occurrence matrix features and growth rates of PH and VIs) statistics for each plant, were extracted. Several selection algorithms (i.e., Boruta, DALEX, genetic algorithm, least absolute shrinkage and selection operator, and recursive feature elimination) were used to select the variable sets useful for predicting SM, FW, and FN. Random forests, ridge regressions, and support vector machines were used to predict the yield using the top five selected variable sets. Results First-order statistics of PH and VIs collected during the early to mid-fruit formation periods, about one month prior to harvest, were important variables for predicting SM. Similar to the case for SM, variables collected approximately one month prior to harvest were important for predicting FW and FN. Furthermore, variables related to PH were unimportant for prediction. Compared with predictions obtained using only first-order statistics, those obtained using the second-order statistics of VIs were more accurate for FW and FN. The prediction accuracy of SM, FW, and FN by models constructed from all variables (rRMSE = 8.8–28.1%) was better than that from first-order statistics (rRMSE = 10.0–50.1%). Conclusions In addition to basic statistics (e.g., average and standard deviation), we derived second-order statistics of PH and VIs at the plant level using the ultra-high resolution UAV images. Our findings indicated that our variable selection method reduced the number variables needed for tomato yield prediction, improving the efficiency of phenotypic data collection and assisting with the selection of high-yield lines within breeding programs.