Multi-encoder ConvNeXt Network with Smooth Attentional Feature Fusion for Multispectral Semantic Segmentation

Researchers have unveiled MeCSAFNet, a new AI architecture for analyzing multispectral satellite imagery. It uses dual encoders to separately process visible and non-visible light channels, then fuses the data with an attention mechanism. The model achieved massive performance gains, beating the popular U-Net model by up to 19.21% in accuracy (mIoU) on the Five-Billion-Pixels dataset. It also outperformed other leading models like SegFormer and DeepLabV3+ on benchmark tests.