IMPROVED SOIL MOISTURE RETRIEVAL USING MACHINE LEARNING (ISML)

Terrestrial land surfaces are characterised by strong heterogeneities of, among other variables, soil texture, orography, land cover, snow, or Soil Moisture (SM). SM is of broad scientific interest due to its role in the Earth system: It impacts the partitioning of the incoming water and energy over land and affects then the variability of the terrestrial water and energy cycle. SM is also of capital practical value for a wide range of applications from floods forecasting to agriculture and water management.

The scientific community has made significant progress in estimating SM from satellite-based Earth Observations (EO). Harmonizing the SM retrievals from active and passive MW measurements from instruments that (i) operate at different wavelengths, polarisations and incidence angles; (ii) have diverging spatial, temporal and radiometric resolutions; and (iii) are hardly ever well collocated in space and time, is a true challenge. For a decade, ESA CCI SM project has released a Climate Data Record (CDR) of daily estimates at a 0.25° resolution, that relies on: 1) a physical-based inversion scheme to retrieve SM from passive MicroWave (MW), 2) a statistical retrieval for active MW, and 3) an a posteriori merging of these two products.

Several sources of improvement can however to be investigated for the retrieval techniques: the synergy of the observations in the retrieval phase and the spatial downscaling based on the data fusion with other satellite observations. These improvements should help improve and generate long-term and high-resolution SM products in particular by exploiting innovative machine learning methods.

In order to close the gap between Earth system research requirements and EO data, we therefore intend to investigate two topics:

1. Improvement of the SM retrieval algorithm based on: an a priori merging of active and passive MW observations, and a machine learning method with a Localization strategy to better take into account local conditions.
2. Spatial downscaling based on: a SM-related index, the data-fusion with other high-resolution observations such as SAR or thermal infrared data, and a deep learning image-based retrieval approach.

We are now at a crossroad of opportunities: on the one hand, AI is becoming one of the most transformative technologies of the 21st century, while on the other hand, European EO capability is delivering a totally unique and comprehensive picture of the planet. This project intends to capitalize on this new context of satellite remote sensing for soil moisture.

The overarching objective of the project is to investigate the feasibility to obtain a Machine-Learning (ML) advanced, high-resolution (1 km, daily, 2010-2022), and consistent Soil Moisture (SM) estimate, over the Euro-Mediterranean region.