KL-Divergence as a Proxy for Plant Growth

Regular crop monitoring is essential for crop quality and health. Breeders and farmers regularly observe crops and measure crop status to characterize crop health, growth, and make subsequent informed management decisions. Several monitoring and observation methods exist, typically based on tried-and-tested approaches walking through the field and recording manual measurements or impressions of the crop's growth stages. While reliable, manual crop scouting is labor-intensive and time-consuming. Although recent advances in automated plant phenotyping can decrease scouting effort, human experts must still view the resulting images or videos to determine the specific stages of crop growth. A reliable method for labelling crop stages from time-lapse imagery could play a significant role in plant phenotyping. In this paper, we propose the use of the Kullback-Liebler Divergence (KLD) on day-averaged time-lapse digital images as a method to identify likely changes in crop stage. The KLD of daily average images offers a simple and theoretically sound method of detecting daily changes in crop growth that is relatively straightforward to compute. We measured the crop growth over daily, 3-day and 7-day periods to identify changes over the growing season with an automated peak detection algorithm. This work demonstrates the feasibility of using KLD as an index for identifying points at which the morphology of a crop has changed sufficiently to warrant a change in stage.