A deep learning model to time-profile plant nutrient uptake in a growth accelerator

Abstract

Synchrony between nutrient supply and plant demand is a key performance indicator of enhanced efficiency fertilisers (EEF’s). This study sought to develop a high throughput technique for rapid, non-destructive plant biomass measurements. In three experiments, a 3D camera mounted on a robotic gantry scanned pots weekly. A deep learning neural network (RandLA-net) was trained with colour point cloud (PCD) data to isolate a single central plant from partly overlapping adjacent plants and infrastructure (overall testing accuracy of 0.93 and mean intersection over union, IOU, of 0.90). Segmented voxels counts were strongly related to above ground dry matter (R2 = 0.87; P < 2.2e-16), and key statistics related to nutrient synchrony, for example inflection point of the logistic curve, were successfully measured. The high-throughput technique allowed rapid evaluation of fertiliser treatment performance and relative nutrient synchrony over time. Optimisation of the approach can be achieved by careful model plant selection, limiting target nutrient applications to less than that required for growth to plateau, and including a maximal productivity reference.