Weed segmentation from multispectral UAV images.
Final year research at the University of Peradeniya: segmenting weeds in broadcast paddy fields from multispectral drone imagery with a U-Net neural network.
The problem
In broadcast-sown paddy fields rice and weeds grow mixed together, so weeds are hard to locate. Mapping them from the air allows targeted treatment instead of blanket spraying.
The approach
Manually annotated a rice and weed dataset from multispectral UAV images in CVAT, then trained a U-Net convolutional neural network in Python with PyTorch to segment weeds pixel by pixel.
The engineering challenge
Building an accurate labelled dataset where rice and weeds look almost identical, and getting the model to use the extra spectral bands.
What this represents
A working segmentation pipeline for multispectral imagery, the same kind of remote sensing analysis used in spatial science.