Coffee Flower Identification Using Binarization Algorithm Based on Convolutional Neural Network for Digital Images.

Area of Science:

• Agricultural remote sensing
• Computer vision
• Machine learning

Background:

• Crop-type identification is crucial for agricultural remote sensing, impacting yield prediction and field management.
• Current satellite and UAV platforms struggle with accurate monitoring of small targets like coffee flowers.
• Ground-based time-lapse imaging offers high spatial-temporal resolution for small-scale plantation monitoring.

Purpose of the Study:

• To enhance coffee flower identification accuracy using time-lapse digital images.
• To develop a robust method for small target monitoring in agricultural settings.
• To evaluate the proposed method against existing machine learning models.

Main Methods:

• A hybrid approach combining the OTSU binarization algorithm and a convolutional neural network (CNN) model.
• Utilizing VGGNet for pre-training and initializing the CNN model.
• Training the CNN with selected positive and negative coffee flower samples from digital images.
• Optimizing boundary information using binarization results after initial CNN extraction.

Main Results:

• The proposed method demonstrates improved coffee flower classification accuracy compared to Support Vector Machine (SVM) and standalone CNN models.
• Optimal performance was achieved under specific conditions: a 52.5° depression angle and soft lighting.
• The method reached a Dice (F1) score of 0.80 and an Intersection over Union (IoU) of 0.67 under optimal conditions.

Conclusions:

• The combined OTSU binarization and CNN approach effectively improves coffee flower identification accuracy from time-lapse images.
• This method provides a viable solution for monitoring small agricultural targets, overcoming limitations of traditional remote sensing platforms.
• Further research can explore variations in imaging conditions and model architectures for broader applicability.