AI-based orchard monitoring at night: Enhancing sustainable fruit production through real-time apple detection

Federal Scientific Agroengineering Center VIM, 109428 Moscow, Russia

^* Corresponding author: alexeykutyrev@gmail.com

Abstract

Accurate recognition, classification and segmentation of apple fruits on tree crowns are of key importance for improving the efficiency of remote monitoring and forecasting of fruit orchard yields at different stages of the production process. The study evaluates the performance of the state-of-the-art convolutional neural network model YOLO11 (You Only Look Once version 11) under artificial lighting conditions at night. Transfer training of the models (n, s, m, l, x) is performed to evaluate their performance. The study highlights the complexities arising from the variability of lighting in natural daytime conditions of industrial gardens, which makes object recognition difficult due to the influence of various natural factors. The results of this research found that night conditions with controlled artificial lighting contribute to improved recognition accuracy. The average accuracy of the models at an IoU of 50% (mAP50) ranges from 0,76 to 0,80, and the mAP50-95 metric ranges from 0,40-0,45. The average absolute error of the models in counting the number of apple tree fruits in the test sample images at night does not exceed 8%. Adaptive learning strategies and ensemble methods can further improve the recognition accuracy under different lighting conditions. Further research is planned to optimize lighting systems to improve the stability of models for real-time operation.