Combining Image Classification and Unmanned Aerial Vehicles to Estimate the State of Explorer Roses

Abstract

The production of Explorer roses has historically been attractive due to the acceptance of the product around the world. This species of roses presents high sensitivity to physical contact and manipulation, creating a challenge to keep the final product quality after cultivation. In this work, we present a system that combines the capabilities of intelligent computer vision and unmanned aerial vehicles (UAVs) to identify the state of roses ready for cultivation. The system uses a deep learning-based approach to estimate Explorer rose crop yields by identifying open and closed rosebuds in the field using videos captured by UAVs. The methodology employs YOLO version 5, along with DeepSORT algorithms and a Kalman filter, to enhance counting precision. The evaluation of the system gave a mean average precision (mAP) of 94.1% on the test dataset, and the rosebud counting results obtained through this technique exhibited a strong correlation (R2 = 0.998) with manual counting. This high accuracy allows one to minimize the manipulation and times used for the tracking and cultivation process.