Weed Classification Using Explainable Multi-Resolution Slot Attention.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Deep neural networks (DNNs) offer potential for weed identification in agriculture but often lack interpretability at high resolutions.
• Existing methods struggle to pinpoint discriminative weed features accurately for effective weed population control.

Purpose of the Study:

• To develop an explainable DNN approach for high-resolution weed identification using attention maps.
• To enhance the interpretability of DNN decisions for agronomists, aiding in the creation of weed identification keys (WIK).

Main Methods:

• A multi-layer attention procedure based on transformers was employed.
• A fusion rule, utilizing a weighted average of saliency-based attention maps from different layers, was developed.
• The model was trained and evaluated on the Plant Seedlings Dataset (PSD) and Open Plant Phenotyping Dataset (OPPD).

Main Results:

• High-resolution attention maps were generated, providing visual explanations for DNN classifications.
• The method achieved high classification accuracies: 95.42% (negative) and 96% (positive) on PSD, and 97.78% (negative) and 97.83% (positive) on OPPD.
• Cross-dataset evaluations highlighted model performance and misclassification insights.

Conclusions:

• The proposed transformer-based attention mechanism significantly enhances the explainability of DNNs for agricultural weed identification.
• High-resolution attention maps provide valuable insights for agronomists, improving weed identification keys and crop management strategies.
• The method demonstrates robust performance across different agricultural datasets, paving the way for advanced precision agriculture techniques.