Tailoring convolutional neural networks for custom botanical data.

Area of Science:

• Agricultural technology
• Computer vision
• Plant pathology

Background:

• Automated disease, weed, and crop classification using computer vision is crucial for future agriculture.
• Existing models like ResNet and EfficientNet often underperform on specialized agricultural datasets.

Purpose of the Study:

• To address the limitations of current models on specialized datasets.
• To develop and evaluate a novel convolutional neural network architecture, PhytNet, for agricultural applications.
• To investigate the spectral characteristics of cocoa trees for informed data collection.

Main Methods:

• Developed a new convolutional neural network architecture named PhytNet.
• Utilized a novel dataset of infrared cocoa tree images.
• Informed data collection using spectroscopy to understand spectral characteristics.
• Compared PhytNet's performance against ResNet and EfficientNet architectures.

Main Results:

• PhytNet demonstrated excellent attention to relevant features with minimal overfitting.
• Existing models like ResNet18 and EfficientNet variants showed signs of overfitting.
• PhytNet achieved an exceptionally low computation cost of 1.19 GFLOPS.
• Spectroscopy data provided insights into informative spectral bands for disease detection.

Conclusions:

• PhytNet is a promising candidate for rapid plant disease classification and precise symptom localization.
• Informative light spectra for cocoa disease detection lie outside the visible spectrum.
• Focusing on local symptoms is more effective for cocoa disease detection than systemic effects.