Supraglacial ponds on debris covered glaciers in High Mountain Asia (HMA) can locally store surface runoff and prolong water delivery to downstream river basins (Benn et al., 2012; Miles et al., 2019). Temporal variation in the extent of supraglacial ponding therefore affects the timing and availability of water supplies to communities downstream. Large supraglacial ponds can rapidly drain during Glacial Lake Outburst Flood (GLOF) events, which presents a hazard risk to local populations (Nie et al., 2018). Furthermore, the rate of glacier surface mass loss is locally enhanced by supraglacial ponds, so that supraglacial ponds influence the sensitive response of HMA glaciers to climate change (Sakai et al., 2000).

Region-wide, multi-temporal maps of supraglacial pond cover are therefore required as inputs to glacier hydrology and mass balance models (Miles et al., 2020). Remote sensing techniques are often used for supraglacial pond mapping because of their wide spatial coverage and capacity for repeat observations. However, current approaches are limited by factors including efficiency and poor transferability throughout time and space (Watson et al., 2018; Wangchuk and Bolch, 2020).

This study develops an efficient, accurate pond mapping approach that is widely spatiotemporally applicable in the HMA region. An unsupervised k-means classifier is used to train a supervised Random Forest Classifier (RFC) in the Google Earth Engine platform. This study adapts algorithms used by Dell et al., (2021) for application in the HMA region. The classifier is trained on four spatially distal glaciers within the Himalaya region, with Sentinel-2 optical satellite imagery obtained from April 2017 to October 2021 and 8m resolution HMA DEM data. The RFC is validated against manually derived pond outlines at these glaciers.

The RFC achieves accuracy of 86.3-99.6% against manually derived outlines for supraglacial ponds with an area greater than 1000m2. A Root Mean Square Error of 978.4m2 is calculated across 222 overlapping pond outlines included in the validation process. The classifier is an important contribution to the existing literature because it is capable of mapping ponds with area less than <10000m2 across multiple glacier valleys and timepoints accurately and efficiently.

The RFC is applied to two of the training sites, Tshojo Glacier, Bhutan Himalaya and Langtang Glacier, Nepal Himalaya, during March-October 2017-2021. This study period encompasses months designated in this study as the Pre-Monsoon (March-April), Early-Monsoon (May-June), Late-Monsoon (July-August) and Post-Monsoon (September-October) periods. At Tshojo Glacier, a seasonal pattern of pond filling during the Pre-Monsoon and Early-Monsoon and drainage in the Late-Monsoon is observed. This pattern is consistent with other literature observations and is indicative of the operation of a rapid englacial drainage mechanism triggered by increased hydraulic potential gradients between ponds throughout the monsoon season (Benn et al., 2012; Miles et al., 2017; Narama et al., 2017).

At Langtang Glacier, a pattern of continual pond expansion in the central ablation zone is observed from 2018-2021. This pattern is unusual and contrasts earlier observations of seasonality at Langtang Glacier (Miles et al., 2017; Steiner et al., 2019). Supraglacial pond seasonality at Langtang Glacier therefore requires extensive further investigation through remote sensing and field observations. However, this study makes the very tentative suggestion that decreasing ice velocity in the central ablation zone may have recently caused compressional ice flow that closes relict englacial conduits, so that they are not available as drainage routes during the Late Monsoon. The findings of this study indicate that complex inter-connected climatic, hydrological and glaciological processes control the evolution of supraglacial ponds in the Himalaya region.