Weed infestation continues to be a significant impediment to sustainable and efficient crop production. Precision agriculture aims to overcome such hurdles through reliable management processes that account for the complexity of site-specific knowledge and data on weeds. UAV-based multispectral imaging can capture bands beyond the visible spectrum including Near-Infrared and Red-Edge, can improve crop–weed differentiation moving from traditional management practices based on traditional RGB images. However, robust semantic segmentation using multispectral data is still challenging due to spectral variation, occlusions, and the heterogeneous nature of field settings.
This thesis proposes a hybrid CNN–Transformer segmentation framework tailored for multispectral crop–weed mapping. The model integrates modality-specific ConvNeXt encoders for spectral feature extraction, Swin Transformer blocks for global contextual reasoning, a gated Feature Pyramid Network (FPN) for adaptive multispectral fusion, and a Pyramid Pooling Module (PPM) for multi-scale decoding.
When evaluated on the WeedsGalore dataset, the proposed model achieved a mean Intersection-over-Union (mIoU) of 90.04%, a considerable improvement over conventional CNN-based and RGB-only baselines. Furthermore, zero-shot and few-shot fine-tuning studies on carrot and onion field datasets show that the proposed model has promising cross-domain generalization ability while learning from limited labeled examples. These findings highlight the potential for multispectral fused learning in conjunction with hybrid architectures to drive site-specific weed management, paving the way towards more scalable and sustainable agricultural practices.
Use Your Cell Phone as a Document Camera in Zoom
From Computer
Log in and start your Zoom session with participants
From Phone
To use your cell phone as a makeshift document camera
