Author(s): Jiahui Qiu; Jari Silander; Kari Luojus; Cemal Melih Tanis; Harri Kaartinen; Yubao Qiu; Epari Ritesh Patro; Bjorn Klove; Ali Torabi Haghighi
Linked Author(s): Björn Klöve
Keywords: River ice; Finland; Camera system; Semantic segmentation; Convolutional neural networks
Abstract: Finnish rivers undergo a pronounced annual freeze-thaw cycle with important ecological and socio-economic implications, yet many Finnish rivers are too small and narrow for reliable monitoring with conventional large-scale remote sensing. Under CRYO-RI and Digital Waters (DIWA) Flagship initiatives, we deployed a shore-fixed camera system along selected sections of the Kiiminkijoki River near its estuary to the Gulf of Bothnia as a demonstration for wider application in Finland. Using oblique-view RGB imagery, we evaluated five semantic segmentation architectures: DeepLabv3+, FPN, UNet++, UNet, and PSPNet, for pixel-level mapping of open water, river ice, and snow cover. Validation against a held-out annotated dataset showed that DeepLabv3+ achieved the best performance, with a mean Intersection over Union (mIoU) of 0.920. The results demonstrate the potential of camera-based deep learning for high-resolution river ice monitoring and provide a transferable framework for future integration with (SYKE) camera networks, Earth-observation validation, and digital-twin-based decision support in Finland.
Year: 2026