Mapping coffee yield with computer vision

Yield maps guide investigations into the causes of spatial and temporal variations in crop yields. The objective of this work was to implement an algorithm based on computer vision to quantify the number of coffee fruits and to build yield maps. Data were collected in two areas of a commercial Arabica coffee (Coffea arabica) plantation. The images of the coffee fruits were taken from 90 videos acquired during the harvest. The You Only Look Once version 4 (YOLOv4) model was used for the detection and counting of coffee fruits. Geographic coordinates were registered at the same time the videos were recorded and associated with video frames. The number of coffee fruit detections for each frame was converted into yield considering the average distance covered by the harvester and the distance between coffee rows of 4 m. The yield maps were interpolated from the video frames' respective geographic coordinates. The YOLOv4 model had a mean average precision of 83.5%. The yield map estimated from the detections obtained by the computer vision model was able to explain 81% of the variance of the reference yield map. The main contributions of the proposed methodology are its low implementation cost and the independence of specific brands of coffee harvesters for the implementation of the image capture structure. Another advantage of this methodology is the possibility of saving raw data from the entire harvest. Thus, its user can pursue further improvements for the model in the future and validate its performance under different scenarios.

Automated summary

This study implemented a computer vision algorithm to quantify the number of coffee fruits and create yield maps. The results showed that this method effectively explained the factors influencing yield variations and had the advantages of low cost and independence from specific coffee harvester brands.