2.1.1. Study Area

UAV RGB imagery has been acquired over three different potato UK fields (P1, P2, P3), one Australian field (P4), and two different UK lettuce fields (L1, L2). Table 1 summarizes the information related to the acquisition of the imagery fore each field. The acquisition was planned in order to cover the full span of growth stages from emergence (plant size of more than 8 cm) to the point where the canopy closes. This translates into a time window for the UAV flights from 4 to 7 weeks after emergence. Having imagery over several dates allows for corregistering our imagery and reuse annotations for all images, making the data annotation process more efficient while capturing data at different growth stages. The imagery for lettuce plants only includes two acquisition dates for two different fields. However, the lettuce fields present a wider range of growth stages, compensating for the reduced number of flight dates and increasing the variety in growth stages in the dataset. Imagery acquisition is carried out on diffuse light days two hours before or after solar noon to avoid shadows and hot-spots. UAV photographs are then stitched up together into a unique orthomosaic image using a Structure-from-Motion algorithm. The annotation process will be considerably facilitated by the use of orthomosaic images. From these images, image tiles are generated with a size of 256 × 256 pixels to be fed into the model. RGB UAV images of potato plants and lettuce plants have been acquired with a resolution between 1.7 cm and 2.0 cm per pixel. At such resolutions, the canopy size and shape can be correctly identified with enough discrete pixels from the soil and vegetation classes. The fields are composed of sandy or silty soils which look similar in UAV RGB imagery. Figure 1 shows a 10 × 10 m sample from each orthomosaic image included in our datasets.
