Orchard Mapping with Deep Learning Semantic Segmentation.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Remote Sensing

Background:

• Accurate orchard tree segmentation is crucial for precision agriculture and yield estimation.
• Existing methods often struggle with variations in seasons, tree age, and environmental factors.

Purpose of the Study:

• To develop and evaluate a deep learning-based approach for automated orchard tree canopy segmentation.
• To assess the model's performance under diverse conditions including different seasons, tree ages, and weed coverage levels.

Main Methods:

• Utilized a U-net convolutional neural network variant for image segmentation.
• Trained and validated the model on a dataset of aerial images from three walnut orchards, encompassing seven use cases.
• Tested the model on unseen orthomosaic images using oversampling and undersampling techniques.

Main Results:

• Achieved high accuracy rates: 91% (training), 90% (validation), and 87% (testing).
• Demonstrated exceptional robustness by reaching up to 99% performance on novel orthomosaic images, despite their absence in the training set.

Conclusions:

• The proposed U-net based approach offers a robust and accurate solution for orchard tree segmentation.
• The method shows significant potential for automated agricultural monitoring and management applications.