A Two-Stage Self-Supervised Learning Framework for Winter Crop-Weed Image Classification.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Precision agriculture demands accurate crop and weed identification, yet annotated data for winter systems is scarce.
• Existing methods struggle with the visual similarities between certain crops and weeds in diverse agricultural settings.

Purpose of the Study:

• To develop and evaluate a two-stage deep learning model for classifying winter crops and weeds.
• To introduce the WinterCropWeedDB, a novel dataset for winter crop and weed image classification.
• To assess the effectiveness of self-supervised pre-training followed by supervised fine-tuning.

Main Methods:

• A novel dataset, WinterCropWeedDB, comprising 1,136 images of six winter crops and four weed species was created.
• An EfficientNet-B3 model underwent self-supervised pre-training using a SimCLR-style approach with InfoNCE loss.
• The pre-trained model was fine-tuned using supervised learning and evaluated on a 30% internal validation split.

Main Results:

• The self-supervised pre-training reduced average contrastive loss from 2.0712 to 1.6835.
• The fine-tuned model achieved a maximum validation accuracy of 98.27% and a macro-F1 score of 0.98.
• Grad-CAM and Grad-CAM++ visualizations confirmed the model's focus on relevant image regions.

Conclusions:

• The integrated self-supervised and supervised deep learning approach is effective for winter crop and weed classification.
• The proposed method demonstrates viability for region-specific agricultural image analysis.
• Further validation on independent datasets is recommended to confirm generalizability.