Novel method for crop growth tracking with deep learning model on an Edge Rail Camera

Yield prediction is an essential part of farm management and has been investigated with various kinds of data and technologies in the last decades. With the advent of deep learning technology, recent studies are focusing on crop growth analysis with image processing. Instead of measuring crops in a destructive way, image analysis enables crop measurement without manipulation of the crop itself. Counting crops using tracker algorithms such as DeepSORT is one of the famous approaches for yield prediction and analysis. However, to enable crop growth monitoring and analysis, it needs consideration of temporal analysis along with spatial analysis. It should be able to compare the previous status of the target crop to the current status to analyze the growth, for example, from bud to flower to strawberry. This paper proposes a novel method for monitoring crop growth with crop clustering. Instead of counting the crops from the images, the proposed methods recognized a crop cluster from the image and measured how it changed during its lifespan. Further, the proposed method is implemented in an edge device for a greenhouse that is able to collect and measure. The proposed method has been validated on a strawberry greenhouse for around a year, which shows MoTA score from 0.57 to 0.86, with respect to the dataset.