EchoNet: Efficient railway segmentation from satellite imagery

Accurate railway segmentation from satellite imagery is the prerequisite of transportation infrastructure mapping and monitoring, yet remains challenging due to the sparse and discontinuous geometry of rail networks, substantial intra-class texture variation, and the stringent precision required by downstream geospatial applications. This study presents EchoNet, a lightweight deep learning framework designed for efficient and accurate railway extraction, together with RailSAT, a new dataset comprising 1,250 satellite image tiles collected across the United States. The proposed architecture features a tailored decoder integrating three key components: a Semantic-guided Implicit Spatial Decoder (SISD) that restores spatial resolution through semantic-driven channel modulation without explicit upsampling; a Lite Contextual Decoder (LCD) employing dilated convolutions to capture long-range structural context; and a Pixel-wise Complementary Attention Fusion (PCAF) module that adaptively merges semantic and spatial features. With only 2.97 M parameters and 0.83 GFLOPs, EchoNet achieves an F1 score and IoU within 0.01 of mainstream baselines, while attaining a throughput of 791 images/s and a peak memory usage of 7 GB on 512 × 512 inputs with a batch size of 200 on a single NVIDIA RTX A6000 GPU. Experimental and ablation results validate the effectiveness of each decoder component. Overall, EchoNet demonstrates high accuracy and computational efficiency, offering a scalable solution for large-scale railway segmentation in satellite remote sensing.