Deep Learning Reveals Soil Moisture-Precipitation Coupling across Europe
Abstract
Soil moisture affects the temperature and humidity profile of the atmosphere, thereby influencing the development and onset of precipitation. However, it remains an open question if an increase in soil moisture leads to an increase or decrease in precipitation. Here, we combine interpretable deep learning and causality principles to address this question. We train a convolutional neural network (CNN) on ERA5 atmospheric reanalysis data to predict precipitation, given soil moisture as one of carefully chosen input variables. Subsequently, a sensitivity analysis of the trained CNN reveals the impact of soil moisture changes on precipitation predictions of the CNN. Because the input variables chosen in addition to soil moisture approximate a sufficient set, an important concept from causality research, this impact approximates the causal impact of soil moisture changes on precipitation. Applying the methodology at a time scale of three hours and a spatial scale of 0.25 degrees across Europe, we find that local increases in soil moisture lead to local increases in precipitation (positive local impact), but to decreases in precipitation in a surrounding region (negative non-local impact). The negative non-local impact tends to exceed the positive local impact resulting in a predominantly negative regional impact. Our methodology and findings contribute to the long-standing debate on the dominant sign of soil moisture-precipitation coupling. Moreover, we expect the demonstrated, novel application of deep learning to lead to further insights into the coupling of variables in complex dynamical systems in the Earth sciences and other scientific fields.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H33F..03T