Sorghum crops classification and segmentation using shifted window transformer neural network and localization based...

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Global population growth necessitates enhanced food security, with sorghum being a key staple crop in developing nations.
• Accurate sorghum yield prediction is crucial for increasing productivity.
• Automated methods for crop analysis show promise but face challenges with sorghum's variable appearance.

Purpose of the Study:

• To develop and evaluate advanced AI models for sorghum crop classification, localization, and segmentation.
• To address challenges in sorghum image analysis caused by variations in color, shape, lighting, and noise.
• To improve the accuracy of sorghum yield estimation through precise crop head analysis.

Main Methods:

• A shifted window transformer (SWT) network was designed for sorghum classification.
• The YOLOv9-c model was employed for sorghum region localization.
• A transformer-based SegNet model, utilizing a fine-tuned SegFormer-B0, was developed for accurate sorghum segmentation.

Main Results:

• The proposed models demonstrated superior performance compared to existing published works.
• The integrated approach achieved accurate classification, localization, and segmentation of sorghum crops.
• The developed methods effectively handle image complexities arising from lighting and noise variations.

Conclusions:

• The study presents a robust AI-driven methodology for sorghum crop analysis.
• The advanced models offer a significant contribution to improving sorghum yield estimation accuracy.
• This research supports efforts to enhance agricultural productivity and ensure global food security.