Deep Learning-Driven Automatic Segmentation of Weeds and Crops in UAV Imagery

Area of Science:

• Precision Agriculture
• Remote Sensing
• Deep Learning
• Computer Vision

Background:

• Traditional weed management is inefficient and environmentally damaging.
• Accurate crop and weed segmentation is crucial for optimizing resource use and yield.
• Unmanned Aerial Vehicle (UAV)-based remote sensing offers a promising alternative for agricultural monitoring.

Purpose of the Study:

• To develop and evaluate a novel precision agriculture framework integrating UAV remote sensing with Super-Resolution Reconstruction (SRR) and semantic segmentation.
• To systematically assess deep learning models, including CNNs, Transformers, and Mambas, for crop-weed segmentation.
• To create and release an annotated dataset for tobacco fields to support future research.

Main Methods:

• Super-Resolution Reconstruction (SRR) was applied to enhance low-quality UAV imagery.
• Semantic segmentation models (CNN, Transformer, Mamba) were employed to differentiate crops from weeds.
• An ensemble model combining Transformer (DPT with DINOv2) and Mamba architectures was developed.

Main Results:

• RCAN achieved optimal SRR performance with a PSNR of 24.98 dB and SSIM of 69.48%.
• The Transformer-Mamba ensemble model yielded the highest mean Intersection over Union (mIoU) of 90.75% for segmentation.
• The framework demonstrated robustness to common environmental disturbances, with an optimal magnification factor of 4×.

Conclusions:

• The integrated SRR and semantic segmentation framework significantly improves crop-weed identification accuracy.
• Transformer and Mamba-based models show superior performance in complex agricultural environments.
• This research provides an efficient and precise tool for crop management, advancing sustainable farming practices.