UAV Imagery Enhanced with Vegetation Indices for Accurate Weed and Tobacco Segmentation Based on Deep Learning

Accurate monitoring of weeds and tobacco in fields can improve yield and reduce costs. This study integrates Unmanned Aerial Vehicle (UAV) imaging with deep learning for tobacco-field monitoring. Video data were captured by a DJI Mini 3 Pro UAV at $1 \mathrm{~m} / \mathrm{s}$ and 10 m altitude. Frames were extracted at 2 fps and overlapped by $70 \%$. An orthomosaic was built using Structure-from-Motion. Cropped patches of $256 \times 256$ pixels covering $25 \%$ of the area formed a dataset of 840 images. All images were split into training, validation, and test sets at an $8: 1$: 1 ratio. ExG, NGRDI, VARI, and RGBVI indices were computed and normalized to the $\mathbf{0}-\mathbf{2 5 5}$ range. Four-channel images were built by adding each index to RGB. An enhanced FCN with $\mathbf{1 0 ~ M}$ parameters and a $\mathbf{3 5}$ M-parameter U-Net were trained. Training used cross-entropy loss with data augmentation via random rotations, flips, brightness changes, and cropping. FCN-4ch-ExG achieved the best performance with $77.7 \% \mathrm{mF}$-score and $69.9 \%$ mIoU on test images. U-Net-3ch obtained $77.8 \% \mathrm{mF}$-score and $70.0 \%$ mIoU, and $78.3 \% \mathrm{mF}$-score and $70.4 \%$ mIoU with ExG. Indices NGRDI and VARI degraded weed detection due to attenuated weed features. Results show adding suitable and simple index preprocessing can boost segmentation without adding model complexity.