Mapping the hydrological system of the Greenland ice sheet in coarse satellite products with a machine learning approach
This paper utilizes supervised machine learning approach to infer presence of water features in Moderate Resolution Imaging Spectroradiometer imagery of Greenland ice sheet in the area UPE_U PROMICE automatic weather station (AWS). Reference dataset consisted of water feature mask was extracted from...
| Main Author: | |
|---|---|
| Format: | Dataset |
| Language: | unknown |
| Published: |
GEUS Dataverse
2020
|
| Subjects: | |
| Online Access: | https://dx.doi.org/10.22008/fk2/ywc9xa https://dataverse01.geus.dk/citation?persistentId=doi:10.22008/FK2/YWC9XA |
| Summary: | This paper utilizes supervised machine learning approach to infer presence of water features in Moderate Resolution Imaging Spectroradiometer imagery of Greenland ice sheet in the area UPE_U PROMICE automatic weather station (AWS). Reference dataset consisted of water feature mask was extracted from high resolution LANDSAT 8 images and co-located with MODIS imagery to fit the Gradient Boosting Decision Tree model. Despite a highly imbalanced dataset and difficult task, model obtained recall of 28.7% and precision of 40.1% for the water feature class and overall F1 Score of 96.5%. Then, the model was used to produce water masks over the period of time from 2000 to 2012. Obtained results showed a positive correlation with the average air temperature in-situ measurements and moderate consistency along entire period with other research studies. The study shows promising results for the use of machine learning for water feature extraction and suggests using different approaches to enhance results. The code accompanying this project is available in the author’s Github repository. |
|---|