Mapping thermokarst lakes and ponds across permafrost landscapes in the Headwater Area of Yellow River on northeastern Qinghai-Tibet Plateau

The large variety of the semi-automated methods in mapping the surface water bodies and the frequent omission of ponds (< 10,000 m(2)) from permafrost regions inventories motivates this work. Based on the correlation matrix, several widely used classification methods for mapping the surface water bodies were assessed. Water bodies inventory was generated at a local and landscape scale in the Headwater Area of the Yellow River (HAYR) on a Sentinel-2 satellite image from 23 November 2015. The assessed methods are: spectral water indices, supervised and unsupervised classifiers (k-means, Density Slicing, Maximum Likelihood Classification - MLC), and machine learning algorithms (Random Forest and Support Vector Machines). The MLC method applied on visible and near-infrared (NIR) bands represents the best ratio regarding the accuracy (96%), Kappa coefficient (0.87), and water surface (14.87 km(2)). However, misclassifications are still present, which requires manual editing. Based on the MLC approach, 651 more water bodies were identified than previous inventories. Ponds are account for up to 93% of the 966 of individual water bodies and contribute to 42% of the total water surface, in the context that were omitted before. This analysis emphasize the importance of method settings for the classifier performance, as well as the ponds abundance and substantial contribution to the total water surface. This inventory will improve the general circulation models and provides baseline information for sustainable management of water resources in the HAYR, one of the core Asian Water Towers.