Abstract
Semantic segmentation of remote sensing images is vital in computer vision for detailed scene understanding, with applications spanning agriculture, environmental monitoring, and urban planning. In this study, we conduct a comparative analysis of advanced RGB image segmentation methods using various Fully Convolutional Neural Network (FCN) architectures. The architectures evaluated include Attention U-Net, Deeplabv3+, U-Nets with backbones such as ResNetv2, MobileNetv2, VGG19, and EfficientNet. Our experiments on UAV-captured images from the Mpokwe Extension Planning Area (EPA) in Zomba, Malawi, reveal that the ResNetv2-based U-Net achieves the highest Intersection over Union (IoU) score of 0.91. This model effectively segments classes like maize, water, trees, bare land, built-up areas, and other vegetation types, facilitating applications in agricultural yield estimation, land use analysis, and environmental monitoring. The MobileNetv2-based U-Net provides a good balance between performance and efficiency, making it suitable for real-time applications. This work demonstrates the potential of leveraging cost-effective RGB sensors to achieve competitive performance in image segmentation, contributing to the Sustainable Development Goals (SDGs) by informing agricultural practices and environmental stewardship.
