Maize Ear Sensing for on-Farm Yield Predictions

In maize (Zea mays L.), early yield prediction is a concept of great interest to breeders and agronomists. Limited studies have tried to leverage field-scale imaging data to perform such crop yield predictions. For this purpose, this study aims to provide a complete data pipeline that uses single-view depth and RGB image data to extract morphological traits of maize ears (length, width, and volume), to forecast yield in a non-destructive approach. A depth camera from a smartphone device was used to acquire imagery data, and those images were used to train the YOLOv12nseg model for segmenting the maize ears. The segmentation masks were then used to cut out the background from the point clouds. These point clouds then underwent a further filtering process to remove the remaining noise. The length and width of the ear were extracted using a geometric computational approach, while the volume was predicted using the deep learning model developed for this project, the Ear Volume Network (EVNet). Lastly, a yield prediction was obtained by using the ear morphological traits as input in the gradient-boosting decision trees. The segmentation task performed well under adverse environmental conditions such as occlusion, noise and variable lighting. The EVNet accurately predicted ear volumes under ideal scenarios ($RMSE =28.91 \text{ml}$). Likewise, the yield forecast step demonstrated high accuracy in both ideal ($R M S E=13.90 g)$ and real (RMSE $=24.12 ~\mathrm{g})$ scenarios. The results of this study highlight the potential of using new technologies to predict yield under field conditions.