| Summary: | As an important component in the terrestrial cryosphere, the soil freeze-thaw (FT) cycle plays a determinant role in climatic, hydrological, ecological, and biogeochemical processes in permafrost landscapes. The FT-state can be monitored exactly with in-situ field measurements, which is costly and limited to single chosen sites. Remote sensing data provides the possibility of collecting information over a large area repeatedly. To explore a more effective way to monitor the FT states in the terrestrial cryosphere, we used microwave and optic remote sensing data and introduced the Deep Learning approach to simulate the soil FT state. MLP and CNN networks were trained and tested with, respectively, over 35000 and about 54000 randomly selected data samples over the entire western part of Nunavik, Canada. The data were labeled following chosen FT reference periods in a year. The trained CNN networks generally performed better than MLP networks and reached model accuracies of around 95%. Different feature combinations and FT references were tested and discussed. The model performance is further validated with a ground truth data set and an extended conceptual ground truth data set. These two sets are built independently of the training and testing sets based on in-situ data from 30 measurement stations from all seasons in a year and reached the accuracies of 87.27% and 94.70, respectively.
|
|---|