A long-term consistent soil moisture dataset based on machine learning and remote sensing
Abstract
Soil moisture is a key climatic variable essential for a wide range of hydrological, vegetation, and energy applications. To understand and monitor global trends in soil moisture and accurately assess water availability and water stress, its vulnerability to climate change and human activity, a long-term soil moisture dataset should have a consistent quality. To tackle this goal, we utilize advanced machine learning to merge multiple source of remote sensing data. We identify a very high-fidelity NASA's Soil Moisture Active Passive (SMAP) soil moisture data as the desired data quality by using it as training labeled data for supervised learning. Then, neural networks are trained to return soil moisture data from brightness temperature data from multiple satellites (SMOS(2010-present), AMSR-E (2002-2011), AMSR2 (2012-present) and SSM-I (1978-present)) going backward in time using the latest optimal synergistic dataset as the optimal one for training in each period. Principle component analysis, branch convolutional neural network architecture, and dense neural network input configuration analysis suggest that information from just one incidence angle is sufficient to reach a high correlation of R = 0.943 against the reference SMAP enhanced Level-3 soil moisture product.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMH002...04S
- Keywords:
-
- 1847 Modeling;
- HYDROLOGY;
- 1848 Monitoring networks;
- HYDROLOGY;
- 1855 Remote sensing;
- HYDROLOGY;
- 1866 Soil moisture;
- HYDROLOGY