Deep Learning Based Apples Counting for Yield Forecast Using Proposed Flying Robotic System.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Convolution Neural Network (CNN) deep learning is prevalent for fruit detection and classification based on visual characteristics.
• Accurate fruit yield forecasting is crucial for producers to establish commercial agreements.

Purpose of the Study:

• To evaluate the performance of two deep learning models, Single-Shot Multibox Detection (SSD) Mobilenet and Faster Region-CNN (Faster R-CNN), for autonomous apple detection and counting.
• To compare the effectiveness of these models when trained on a custom dataset versus pre-trained COCO datasets.
• To enable accurate yield predictions for apple producers through automated orchard monitoring.

Main Methods:

• Generated a custom dataset of 4000 red apple images for training.
• Trained SSD-Mobilenet and Faster R-CNN models using the custom dataset with a learning rate between 0.015-0.04.
• Developed a Flying Robotic System (FRS) for autonomous apple detection and counting in an orchard.
• Experimentally compared the performance of the trained models.

Main Results:

• Both SSD-Mobilenet and Faster R-CNN models achieved accuracy rates up to 93%.
• The Faster R-CNN model demonstrated highly successful detection performance, reducing the loss value below 0.1.
• Autonomous detection and counting were successfully implemented in a commercial apple orchard.

Conclusions:

• Faster R-CNN and SSD-Mobilenet are effective deep learning architectures for autonomous apple detection and yield estimation.
• Training with a custom dataset significantly enhances model performance for specific apple varieties.
• Automated systems utilizing these models can provide valuable tools for agricultural producers to improve yield forecasting.