A New Dataset and Comparative Study for Aphid Cluster Detection and Segmentation in Sorghum Fields

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Aphid infestations cause significant crop damage and yield loss in wheat and sorghum.
• Current pesticide application is inefficient and harmful, leading to waste and environmental concerns.
• There is a critical need for intelligent systems to detect and manage aphid infestations precisely.

Purpose of the Study:

• To develop and evaluate machine learning models for detecting and segmenting aphid clusters in agricultural fields.
• To create a comprehensive dataset for training and testing these models.
• To compare the effectiveness of semantic segmentation and object detection for aphid infestation assessment.

Main Methods:

• A large-scale dataset of 54,742 annotated image patches from sorghum fields was created.
• Four real-time semantic segmentation models and three object detection models were trained and evaluated.
• Performance was assessed based on precision, recall, mAP, and FPS.

Main Results:

• Fast-SCNN achieved 80.46% mean precision and 81.21% mean recall for segmentation with 91.66 FPS.
• RT-DETR showed the best object detection performance with 61.63% mAP and 92.6% mean recall.
• Aphid cluster segmentation was found to be more suitable than detection for infestation assessment.

Conclusions:

• Semantic segmentation models, especially Fast-SCNN, offer a more effective solution for identifying and managing aphid infestations.
• The developed dataset and models support the development of autonomous, targeted pest management systems.
• This approach can reduce chemical pesticide use and mitigate agricultural yield losses.