A Novel Shape Based Plant Growth Prediction Algorithm Using Deep Learning and Spatial Transformation

Plant growth prediction is challenging, as the growth rate varies depending on environmental factors. It is an essential task for efficient cultivation in controlled environments, such as in plant factories. In this paper, we propose a novel deep learning network for predicting future plant images from a number of past and current images. In particular, our focus is on the estimation of leaf shape in a plant, because the amount of plant growth is commonly quantified based on the leaf area. A spatial transform is applied to a sequence of plant images within the network, and the growth behavior is measured using a set of affine transform parameters. Instead of conventional sequential image fusion, the affine transform parameters for all pairs of successive images are fused together to predict the shape of the future plant image. Then, an RGB reconstruction subnet divides the plants into multiple patches to make global and local growth predictions based on hierarchical auto-encoders. A variety of experimental results show that the proposed network is robust to dynamic plant movements and can accurately predict the shapes of future plant images.