Abstract

The continuous growth of the world's population requires immediate action to ensure food security. Sorghum is among the five most-produced cereals and is a dietary staple in many developing countries. Therefore, it is of great importance to obtain precise information for improving cereal productivity. An indicator for estimating sorghum yields is the number of crop heads in different branching arrangements. Approaches based on image processing and artificial intelligence have proved useful for automatically and efficiently obtaining this type of information for different crops. However, their application to sorghum crops presents some additional challenges owing to differences in the shape and color of sorghum heads. In this study, a methodology to detect sorghum heads in unmanned aerial vehicle imagery was investigated, and its performance was evaluated using a standard quality index in object detection problems (mean average precision). Specifically, test-time-augmentation (TTA) techniques have been implemented using a set of geometrical and color transformations selected according to the sorghum plant imagery requiring analysis, as well as four different ensemble learning methods. Because these methods are weighted, two different approaches for calculating these weights to improve sorghum head detection have been proposed. The results show that in sorghum head detection, TTA strategies outperform detection based only on individual transformed testing sets. Moreover, these results were improved by the use of different weights during the ensemble of TTA results.