Development of a machine learning plant growth estimator for hydroponics

Abstract

Growth stage estimation in indoor farming is crucial for precision agriculture. It reduces the wastage of resources such as nutrients and water in plant cultivation. Overly, yield and quality of the agricultural produce are improved while minimising running costs. However, plants at different growth stages pose similar morphological shapes, making growth stage estimation complex. Therefore, an algorithm that can investigate other features of plant leaf besides shape is needed to distinguish plant growth stages accurately. In this work, we suggested three approaches: first, a machine learning approach that combines morphological operators to generate a contour mask dataset that highlights the morphological leaf structures, Gabor filters for feature extraction, and traditional machine learning algorithms for classification, second, a deep learning approach through transfer learning, and third, an approach that uses a feature extractor based on deep learning techniques and a classifier based on traditional machine learning techniques. Side-by-side experiments are run to evaluate the performance of the three proposed approaches In addition, two datasets (canola and radish) were created from a publicly available dataset as "bccr-segset" to develop and test these approaches. We found that the third approach of using a feature extractor based on deep learning techniques and a classifier based on traditional machine learning techniques can classify four plant growth stages with the highest accuracy of 98.4% and a faster average classification time per image of 0.07 seconds. On the other hand, the first approach, the machine learning method, achieved slightly lower classification accuracy than the deep learning models but had the shortest classification time per image of 0.04 seconds.

Finally, we compared our results with similar research studies and found out that our proposed approaches compete well with these studies.