Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Sensor-invariant Satellite Embedding Learning
Design a deep learning approach which integrates the measurements from different types of satellite data (MODIS, VIIRS, Sentinel-1 SAR, Sentinel-2, see Fig. 1) by learning a sensor-invariant embedding (see Fig. 2), targeting two applications: lake ice monitoring and glacial lakes mapping.
Lake ice is an important variable to understand the regional and global climate change and has been recently recognized as an Essential Climate Variable (ECV). Monitoring and analyzing the (decreasing) trends in lake freezing using artificial intelligence techniques provides important information for climate research.
Across the globe, as one of the repercussions of climate change and global warming, several new glacial lakes have formed in the previously glaciated areas. In addition, the area of many existing glacial lakes is on the rise. Prior research showed that rapid deglaciation and lake formation have dramatic effects on downstream ecosystem services, hydropower production and high-alpine hazard assessments. Hence, it is essential to map and monitor the changes in water extent in these lakes at a higher frequency for hazard assessment and Glacial Lake Outburst Flood (GLOF) risk estimation.
This work is part of the following ongoing projects:
- Alpine Regional Initiative (AlpGlacier, funded by ESA)
- GLOFCA project (funded by UNESCO)
A good starting point would be the following work: https://arxiv.org/abs/2107.09092
Supervisors: Dr. Manu Tom, Prof. Dr. Konrad Schindler
Lake ice is an important variable to understand the regional and global climate change and has been recently recognized as an Essential Climate Variable (ECV). Monitoring and analyzing the (decreasing) trends in lake freezing using artificial intelligence techniques provides important information for climate research.
Across the globe, as one of the repercussions of climate change and global warming, several new glacial lakes have formed in the previously glaciated areas. In addition, the area of many existing glacial lakes is on the rise. Prior research showed that rapid deglaciation and lake formation have dramatic effects on downstream ecosystem services, hydropower production and high-alpine hazard assessments. Hence, it is essential to map and monitor the changes in water extent in these lakes at a higher frequency for hazard assessment and Glacial Lake Outburst Flood (GLOF) risk estimation.
This work is part of the following ongoing projects: - Alpine Regional Initiative (AlpGlacier, funded by ESA) - GLOFCA project (funded by UNESCO)
A good starting point would be the following work: https://arxiv.org/abs/2107.09092
Supervisors: Dr. Manu Tom, Prof. Dr. Konrad Schindler
The main tasks are:
- Get acquainted with the existing joint (MODIS, VIIRS, Sentinel-1 SAR) satellite embedding learning approach.
- Make the existing embedding sensor-invaraiant (see Fig. 2). For e.g. by including a contrastive loss term.
- Re-design the current network architecture in an end-to-end fashion and simplify the current multi-step procedure to train the network.
- Using the newly learnt embedding, perform a case-study for lake ice monitoring (3 class categories: frozen lake pixels, non-frozen lake pixels, background pixels):
a) Test the new sensor-invariant embedding using data from four satellites (MODIS, VIIRS, Sentinel-1 SAR, Sentinel-2), four lakes (Sihl, Sils, Silvaplana, St. Moritz), and two winters (2016-17, 2017-18).
b) Integrated monitoring of lake ice using the developed embedding, which also includes the precise determination of ice-on and ice-off dates.
- Using the newly learnt embedding, perform a case-study for glacial lakes mapping (2 class categories: lake, background):
a) Test the new sensor-invariant embedding using data from two satellites (Sentinel-1 SAR, Sentinel-2), for 400+ small glacial lakes from the Swiss Alps, focussing on the data from years 2014-2016
The main tasks are:
- Get acquainted with the existing joint (MODIS, VIIRS, Sentinel-1 SAR) satellite embedding learning approach.
- Make the existing embedding sensor-invaraiant (see Fig. 2). For e.g. by including a contrastive loss term.
- Re-design the current network architecture in an end-to-end fashion and simplify the current multi-step procedure to train the network.
- Using the newly learnt embedding, perform a case-study for lake ice monitoring (3 class categories: frozen lake pixels, non-frozen lake pixels, background pixels): a) Test the new sensor-invariant embedding using data from four satellites (MODIS, VIIRS, Sentinel-1 SAR, Sentinel-2), four lakes (Sihl, Sils, Silvaplana, St. Moritz), and two winters (2016-17, 2017-18). b) Integrated monitoring of lake ice using the developed embedding, which also includes the precise determination of ice-on and ice-off dates.
- Using the newly learnt embedding, perform a case-study for glacial lakes mapping (2 class categories: lake, background): a) Test the new sensor-invariant embedding using data from two satellites (Sentinel-1 SAR, Sentinel-2), for 400+ small glacial lakes from the Swiss Alps, focussing on the data from years 2014-2016