Non-Destructive Biomass and Relative Growth Rate Estimation in Aeroponic Agriculture using Machine Learning

Abstract: Optimising plant growth in a controlled climate requires good measurements of both biomass (measured in grams) and relative growth rate (measured in grams of growth per day and gram of plant). In order to do this efficiently and continuously on an individual level during plant development, this has to be done non-destructively and without frequent and labor intensive weighing of plant biomass. This thesis compares the ability of two machine learning methods, Multi-Variate Regression and Neural Networks, to estimate the biomass and relative growth rate from images of plants. The plant data set consists of images of 57 plants from two angles taken on 1-hour intervals during a 5 day period. The results show that images taken from a top-down perspective are best used with multi-variate regression, while images taken from the side are better when used with neural networks. In addition, using images from both cameras improved the biomass estimates from the neural network, but not those from the multi-variate regression. The predictions were improved in all cases when a moving average was taken of consecutive predictions, which likely reduced short-time variance in the data set. For both methods, the relative growth rate estimates were greatly improved by using estimates from both cameras. The low number of individual plants and high image capture frequency created a lot of correlation within the training set, which likely decreased generalization and lowered the accuracy of the predictions on the test set. The best biomass estimates were made using multi-variate regression with images from the top camera and a moving average filter, resulting in an RMSE of 0.0391 g. This corresponds to a relative RMSE of around 11% which is comparable to previous studies. The relative growth rate estimates were not very accurate, but the best method used a neural network with both cameras, resulting in an RMSE of 0.1767 g/(g·day). This corresponds to a relative RMSE of over 100%. A bigger data set with measurements from a larger set of individual plants during a longer time interval within the cultivation period would likely improve these estimates.