A deep learning-based method for mapping alpine intermittent rivers and ephemeral streams of the Tibetan Plateau from Sentinel-1 time series and DEMs

Flow regime changes of Intermittent Rivers and Ephemeral Streams (IRES) can serve as an indicator under global warming, yet the distribution of IRES is rarely extracted at the alpine catchment scale due to their narrow water surface and the heavy cloud contamination on the Tibetan Plateau. Here, a new two-stage method using the deep learning model is proposed for Mapping alpine IRES from Sentinel-1 time series and Digital elevation models (MISD). Firstly, the median images of cross-orbits double-periods (i.e., the flowing period and the drying-up period of alpine IRES) Sentinel-1 time series are input to the deep learning model to synoptically extract alpine IRES in mixed pixels under the heavy cloud contamination. Secondly, the deep learning-based output is corrected by the critical drainage accumulation derived from digital elevation models to remove the disturbance of the non-channelized overland flow on upland. The MISD method was first applied to an alpine catchment, namely the Duodigou Catchment, and then was assessed in the whole Lhasa River Basin. The results showed that the application of cross-orbits double-periods Sentinel-1 time series in the deep learning model is helpful in handling the mixed-pixel problem. And the critical drainage accumulation correction further improves the deep learning-based output with F2 (the metric that measures precision and recall of model) and the median Euclidean distance error of 0.72 and 64.0 m, respectively. Subsequently, the newly proposed MISD method outperforms other river extraction methods in alpine IRES mapping with higher F2 (increased by 0.5) and lower median Euclidean distance error (decreased by 145.7 m). Moreover, the MISD method is characterized by the capability for detecting narrower or lower flow IRES with river width > 1.7 m and discharge >2 L/s, which is significantly ignored in the current global water products. Furthermore, the MISD method successfully recognizes most of the channelized IRES from permanent rivers in the whole Lhasa River Basin through essential training. Therefore, the MISD method is a powerful tool for monitoring the changes of IRES induced by the glacier retreats or the permafrost degradation, etc., on the warming Tibetan Plateau.

Automated summary

A new two-stage method based on deep learning is proposed to extract alpine IRES from Sentinel-1 time series and digital elevation models. The method utilizes cross-orbits double-periods images and critical drainage accumulation to address mixed-pixel problem and non-channelized overland flow disturbance. The results show that the method outperforms other river extraction methods and can detect narrower or lower flow IRES.