Glacial Lakes Mapping Using Multi Satellite PlanetScope Imagery and Deep Learning

Abstract

Glacial lakes mapping using satellite remote sensing data are important for studying the effects of climate change as well as for the mitigation and risk assessment of a Glacial Lake Outburst Flood (GLOF). The 3U cubesat constellation of Planet Labs offers the capability of imaging the whole Earth landmass everyday at 3–4 m spatial resolution. The higher spatial, as well as temporal resolution of PlanetScope imagery in comparison with Landsat-8 and Sentinel-2, makes it a valuable data source for monitoring the glacial lakes. Therefore, this paper explores the potential of the PlanetScope imagery for glacial lakes mapping with a focus on the Hindu Kush, Karakoram and Himalaya (HKKH) region. Though the revisit time of the PlanetScope imagery is short, courtesy of 130+ small satellites, this imagery contains only four bands and the imaging sensors in these small satellites exhibit varying spectral responses as well as lower dynamic range. Furthermore, the presence of cast shadows in the mountainous regions and varying spectral signature of the water pixels due to differences in composition, turbidity and depth makes it challenging to automatically and reliably extract surface water in PlanetScope imagery. Keeping in view these challenges, this work uses state of the art deep learning models for pixel-wise classification of PlanetScope imagery into the water and background pixels and compares the results with Random Forest and Support Vector Machine classifiers. The deep learning model is based on the popular U-Net architecture. We evaluate U-Net architecture similar to the original U-Net as well as a U-Net with a pre-trained EfficientNet backbone. In order to train the deep neural network, ground truth data are generated by manual digitization of the surface water in PlanetScope imagery with the aid of Very High Resolution Satellite (VHRS) imagery. The created dataset consists of more than 5000 water bodies having an area of approx. 71km2 in eight different sites in the HKKH region. The evaluation of the test data show that the U-Net with EfficientNet backbone achieved the highest F1 Score of 0.936. A visual comparison with the existing glacial lake inventories is then performed over the Baltoro glacier in the Karakoram range. The results show that the deep learning model detected significantly more lakes than the existing inventories, which have been derived from Landsat OLI imagery. The trained model is further evaluated on the time series PlanetScope imagery of two glacial lakes, which have resulted in an outburst flood. The output of the U-Net is also compared with the GLakeMap data. The results show that the higher spatial and temporal resolution of PlanetScope imagery is a significant advantage in the context of glacial lakes mapping and monitoring.

Seasonal Evolution of Supraglacial Lakes on Baltoro Glacier From 2016 to 2020

“…The second study mapping supraglacial lakes on Baltoro Glacier used a U-Net model with EfficientNet backbone based on PlanetScope data (Qayyum et al, 2020). Qayyum et al (2020) classified lake areas of 2.25 km 2 on July 14, 2017, 1.96 km 2 on July 27, 2017, 1.83 km 2 on August 7, 2017, 2.22 km 2 on July 14 and 15, 2018, and 2.61 km 2 on July 12-16, 2019. For complete coverage of the glacier surface in 2018 and 2019, they used PlanetScope data acquired within a period of 2 or 4 days.…”

“…The differences in the classification results can most likely be accounted for by the different ground samplings. In particular, the data of the multisensor time series in our study were sampled to 10 m, whereas the study of Qayyum et al (2020) was based on the ground sampling of 3 m.…”

Comparing Methods for Segmenting Supra-Glacial Lakes and Surface Features in the Mount Everest Region of the Himalayas Using Chinese GaoFen-3 SAR Images

“…The quality and quantity of training data are highly important [34,35] In general, the use of satellite images for the monitoring of glacial lakes has been mostly concerned with conventional manual digitization, edge detection, image segmentation methods, and object-oriented classification, rarely adopting some advanced deep-learning methods. Compared with other methods, deep-learning algorithms have a high ability in terms of feature extraction and autonomous learning; they possess a large number of hidden layers and can support a higher level of data abstraction and prediction.…”

“…They used the pre-trained EfficientNet as the backbone of the U-Net architecture, which can automatically extract the majority of the supra-glacial lakes that were missed in other inventories, but they only observed the inter-annual variations rather than seasonal changes in the glacial lake area. The quantitative evaluation measurements were also not given [35]. Furthermore, in terms of monitoring targets, there is a dearth of research into the extraction of dynamic changes of supra-glacial lake outlines, ice crevasses, and supra-glacial streams in mountainous glaciated areas.…”

ACFNet: A Feature Fusion Network for Glacial Lake Extraction Based on Optical and Synthetic Aperture Radar Images

“…However, to the best of our knowledge, there are few studies on DL applications for glacial lake extraction. Qayyum et al applied a U-Net model to glacial lake extraction on very high resolution PlanetScope imagery and obtained better results than those acquired with SVM and RF classifiers [44]. Chen applied U-Net on supra-glacial lake extraction using high-resolution GaoFen-3 SAR images [45].…”