A CBAM-Enhanced UNetFormer for Semantic Segmentation of Wheat Yellow-Rust Disease Using Multispectral Remote Sensing Images

This study focuses on the problem of wheat yellow-rust disease caused by climate change and incorrect farming methods. Early detection of the disease, which manifests as yellow-orange spores on wheat leaves, is crucial for mitigating issues such as reduced crop yield, increased pesticide use, and environmental harm. Current CNN-based semantic segmentation models focus mainly on processing local pixels, which can be insufficient for large areas. This study proposes a novel version of the UNetFormer architecture, enhancing the CNN-based encoder with CBAM modules while utilizing a Transformer-based decoder to address the limitations of current approaches. Specifically, the model incorporates a Convolutional Block Attention Module (CBAM) to refine feature extraction along spatial and channel axes. CBAM modules allow the network to prioritize meaningful features, particularly near-infrared (NIR) wavelength reflections critical for detecting wheat yellow-rust. The proposed UNetFormer2 model effectively captures long-range dependencies in multispectral remote sensing images to improve disease detection across large agricultural areas. Specifically, the model achieves an IoU improvement of 2.1% for RGB, 4.6% for NDVI, and 3% for NIR compared to the baseline UNetFormer model. This work aims to improve wheat yellow-rust disease monitoring efficiency and contribute to more sustainable agricultural practices by reducing unnecessary pesticide application.