Towards deep learning based smart farming for intelligent weeds management in crops

Area of Science:

• Agricultural Technology
• Computer Vision
• Deep Learning

Background:

• Weed management is crucial for crop production, and traditional methods are labor-intensive.
• Deep learning (DL) offers advanced solutions for automated object detection in various applications, including agriculture.

Purpose of the Study:

• To propose a DL-based model for efficient weed detection in crops.
• To evaluate the performance of different You Only Look Once (YOLO) models for weed identification.

Main Methods:

• Utilized a Convolutional Neural Network (CNN) based object detection system, You Only Look Once (YOLO).
• Trained four YOLO models (v3, v3-tiny, v4, v4-tiny) on a dataset of four weed species.
• Applied LAB and HSV image transformation techniques to the dataset.

Main Results:

• The YOLOv4 model demonstrated the highest accuracy, correctly predicting 98.88% of weeds.
• Achieved an average loss of 1.8 and 73.1% mean average precision.
• Image transformations (LAB, HSV) had minimal impact on model performance.

Conclusions:

• The proposed DL model, particularly YOLOv4, is highly effective for weed detection.
• This technology can significantly improve weed management strategies in agriculture.
• Future integration with variable rate sprayers aims for real-time, precise weed control.