RGB-Based Deep Learning for Freeze Damage Detection in Strawberry: Comparing Scratch and Transfer Learning Approaches on Custom Data.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Freeze damage significantly impacts strawberry crops, causing economic losses.
• Manual damage assessment is inefficient, subjective, and labor-intensive.

Purpose of the Study:

• To develop and evaluate automated freeze damage classification for strawberry plants using deep learning and computer vision.
• To compare the performance of different convolutional neural network (CNN) architectures and training methods.

Main Methods:

• Utilized RGB images of strawberry plants for freeze damage classification.
• Evaluated four CNN architectures: DenseNet-121, Inception V3, ResNet-50, and Xception.
• Compared transfer learning (TL) with training models from scratch.

Main Results:

• Models trained from scratch surpassed TL models, with ResNet-50 achieving 97% accuracy.
• ResNet-50 offered the fastest inference time (3.0s), while DenseNet-121 was the most memory-efficient (26.86MB).
• Models effectively identified severe damage but struggled with mild or minimal damage differentiation.

Conclusions:

• Scratch-trained deep learning models provide a more accurate automated solution for strawberry freeze damage classification.
• ResNet-50 is suitable for speed-critical applications, and DenseNet-121 for memory-constrained environments.
• Deep learning offers a rapid, accurate, and nondestructive alternative to traditional freeze damage assessment methods.