A Feedforward Neural Network for ESA CCI Soil Moisture Disaggregation

This study presents a methodology for disaggregating the ESA Climate Change Initiative (CCI) Soil Moisture (SM) maps from 0.25° to a 60 m grid, using a feedforward neural network. This technique is applied over an area of 66,700 km
²
, encompassing parts of Oklahoma and Kansas (US), throughout 2021. The disaggregation approach leverages synergies between different variables, including various Sentinel-2 bands and indices, land surface temperature from MODIS, accumulated precipitation from ERA5-Land, terrain elevation and slope from the STRM, and soil composition. The methodology employs a two-step process: (i) the model is first trained using all the variables at low resolution (0.25°), and (ii) it is then employed to estimate the SM at high resolution using the input variables at 60 m. Results indicate that the model trained at low resolution achieves a reasonably high accuracy over the testing data (RMSE<0.02 m
³
•m
^-3
and R
²
>0.9). The preliminary analysis of the 60 m resolution SM maps and their comparison with two in-situ stations show a strong correlation (R>0.8), and uRMSE close to 0.04 m
³
•m
^-3
, with a bias ranging from 0.022 m
³
•m
^-3
to 0.06 m
³
•m
^-3
.