Fine-Scale River-Ice-Type Mapping During the Freeze-Up Period Using UAV Multispectral Imagery

The spatial distribution of river ice types during the freeze-up period provides important information on channel freezing and ice-cover development and serves as a basis for ice-condition monitoring and ice-jam flood risk assessment. However, river ice types differ considerably in spatial scale, spectral response, and surface structure, making their fine-scale identification under complex ice conditions challenging. In this study, nine representative frozen reaches along the Inner Mongolia section of the Yellow River were investigated using UAV multispectral orthomosaics with a spatial resolution of approximately 0.15 m. Four representative river-surface classes—open water, transparent ice, smooth ice, and rough ice—were mapped. Their responses in the visible, red-edge, and near-infrared bands were analyzed together with spectral indices and gray-level cooccurrence matrix (GLCM) texture features to characterize interclass spectral and textural differences. A deep-learning mapping model integrating multispectral and textural information was subsequently developed and evaluated on spatially independent reaches. Permutation importance was also used to quantify the contribution of each remote-sensing feature. The results showed that the four surface classes exhibited distinct spectral and textural characteristics. Transparent ice and open water showed limited spectral differences in several bands, whereas texture features captured differences in their local spatial structures and improved the model’s ability to distinguish spectrally similar classes. The model combining spectral indices and GLCM texture features accurately delineated the spatial distributions of the river ice types. In the spatially independent test reaches, it achieved an overall accuracy of 95.80%, a Macro-F1 score of 94.53%, and a mean intersection over union (mIoU) of 90.14%. The model achieved an mIoU of 84.55% on an additional independent test reach with complex ice conditions, which was excluded from training and parameter optimization. Feature-contribution analysis identified spectral indices and texture features as important sources of discriminatory information. NDUWI and NDFSI, together with texture measures calculated from the red and near-infrared bands, ranked among the most informative inputs. This result suggests that the four river-surface classes were distinguished not only by their spectral differences but also by variations in local surface patterns. The UAV imagery therefore supported detailed mapping of heterogeneous frozen reaches and provided spatially explicit information relevant to ice-condition monitoring, analysis of river-ice development, and assessment of potential ice-jam flooding.