Multi-encoder ConvNeXt Network with Smooth Attentional Feature Fusion for Multispectral Semantic Segmentation
MeCSAFNet is a dual-branch encoder-decoder designed for land cover segmentation using multispectral imagery. It separates processing streams for visible and non-visible bands using ConvNeXt encoders, then fuses features through a pyramid approach enhanced by CBAM attention and smooth activation functions. Results show superior performance on challenging datasets like Potsdam compared to traditional models.