EdgeSpill-Net: a computationally efficient CNN for semantic segmentation of oil spills in fluvial environment

Oil spills in rivers and canals demand timely, on-site detection, yet most monitoring focuses on maritime incidents and satellite imagery, which lack the spatial/temporal resolution needed for narrow, turbulent waterways. This study presents EdgeSpill-Net, a compact segmentation network tailored to fluvial scenes. The architecture combines a MobileNetV3 backbone with CBAM attention, a five-branch ASPP for multi-scale context, and a lightweight decoder, and is trained with a confidence-guided semi-supervised regimen that exploits a small unlabeled pool while preserving label quality. A task-specific oil-focused loss increases sensitivity to sparse, low-contrast spill pixels. On a held-out riverine test set (three classes: oil, water, other), EdgeSpill-Net attains oil-class F1 = 0.97 and mIoU = 0.926 with only ~2.1M parameters and ~0.6 GFLOPs at 256×256 input. Deployed with ONNX Runtime and INT8 post-training quantization on Raspberry Pi 5 (8 GB) with Sony IMX500 input, the system sustains 2–3 FPS end-to-end with < 400 MB memory, enabling online, on-device river monitoring. These results indicate a favourable accuracy–efficiency trade-off for edge environmental sensing and support early detection and mitigation of inland oil spills.