Estimation of spinach (Spinacia oleracea) seed yield with 2D UAV data and deep learning

Precision agriculture has drawn much attention in the last few years because of the benefits it has on reducing farming costs while maximizing the harvest obtained. Yield prediction is of importance for farmers to fertilize accordingly to reach the potential yield. However, this task is still relying on manual work, which is expensive and time-consuming. Instance segmentation has been implemented in the last years for fruit detection and yield estimation, obtaining state-of-the-art metrics, and reducing the labor required. This research presents a novel approach for spinach seed yield estimation for seed production purposes, that consists of correlating the number of plants and two phenotyping variables (plant area and canopy cover percentage) with the number of harvested seeds and the thousand seed weight. Mask R-CNN is applied to count the number of detections of spinach plants and obtain the object mask from which the plant area is derived. The results show that there is a high linear correlation between a multivariate linear mixed model of the three variables and the number of seeds, with an R²_adj of 0.80. Furthermore, 77.42% of the variation in the weight of thousand seeds can be explained by the number of plants. For future studies, the algorithm should be trained with more spinach images from different locations and under varying weather conditions to allow it to generalize for the crop worldwide. It can be concluded, until further research, that Mask R-CNN can be applied for spinach counting and the computation of its individual plant area, with promising results.