Effectiveness of Training with Procedurally Generated Synthetic Images of Crop Plants

Artificial neural networks are often used to identify features of crop plants. However, training their models requires many annotated images, which can be expensive and time-consuming to acquire. Procedural models of plants, such as those developed with Lindenmayer-systems (L-systems) can be used to create visually realistic images, where annotations are implicitly known. These synthetic images can either augment or be used on their own instead of using real images in training neural networks for phenotyping tasks. The objectives of this paper are two fold. Firstly, we explore the degree to which realism in the synthetic images improves prediction. Secondly, we systematically vary amounts of real images used for training in both maize and canola to better understand situations where only synthetic images generated from L-systems can accurately predict phenotypic properties on real images. We achieved a mean absolute error (MAE) of 1.05 in predicting leaf count in maize and of 1.59 in predicting inflorescence branch count in canola using only synthetic images with U-Net. Furthermore, predictions made with only synthetic images as training data were improved by almost a ten-fold factor (in terms of MAE) by carefully calibrating the procedural model to real images.