% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@ARTICLE{Sivaprasad:1041302,
author = {Sivaprasad, Visakh and Rahmati, Mehdi and Springer, Anne
and Vereecken, Harry and Montzka, Carsten},
title = {{D}evelopment of {C}ontinuous {AMSR}-{E}/2 {S}oil
{M}oisture {T}ime {S}eries by {H}ybrid {D}eep {L}earning
{M}odel ({C}onv{LSTM}2{D} and {C}onv2{D}) and {T}ransfer
{L}earning for {R}eanalyses},
journal = {IEEE journal of selected topics in applied earth
observations and remote sensing},
volume = {0},
issn = {1939-1404},
address = {New York, NY},
publisher = {IEEE},
reportid = {FZJ-2025-02209},
pages = {1 - 16},
year = {2025},
abstract = {Surface soil Moisture (SSM) is a crucial climate variable
of the Earth system that regulates water and energy
exchanges between the land and atmosphere, directly
influencing hydrological, biogeochemical, and energy cycles.
However, satellite-derived SSM, particularly from the
Advanced Microwave Scanning Radiometer AMSR-E/2, is limited
by radio frequency interference (RFI), vegetation effects,
frozen ground, and significant spatial and temporal data
gaps. By excluding data points affected by these problems,
we are able to train an unaffected system and fill the gaps
with high accuracy predictions. We developed a sophisticated
deep learning ConvLSTM model, that combines Convolutional
Long ShortTerm Memory (ConvLSTM2D) layers and Convolutional
Neural Network (Conv2D) layers. The model initially enhances
AMSR-2 SSM values across time and space using Advanced
SCATterometer (ASCAT) SSM as input. The ConvLSTM model,
trained to enhance AMSR-2 SSM, is then fine-tuned by using
the transfer learning technique to enhance AMSR-E data. The
enhanced AMSR-2 data is used as a target to guide the
enhancement of AMSR-E. This approach ensures that gaps in
AMSR-E data are filled, while aligning the characteristics
with the more consistent AMSR-2 SSM, resulting in a seamless
AMSR-E/2 dataset from 2003 to 2023. Unlike previous studies
incorporating additional datasets like precipitation,
temperature, and Digital Elevation Models, our approach
avoids these to prevent redundancy and potential
inaccuracies when generating land surface reanalyses based
on data assimilation, since such data are already integrated
into the land surface model. The ConvLSTM model achieved a
lower RMSE of 0.07 for AMSR-2 prediction and 0.04 for AMSR-E
via transfer learning demonstrating significant gap-filling
accuracy. The enhanced SSM demonstrated a $26\%$ improvement
in the correlation with in-situ SSM measurements, while
maintaining accuracy and consistency in spatial and temporal
patterns.},
cin = {IBG-3},
ddc = {520},
cid = {I:(DE-Juel1)IBG-3-20101118},
pnm = {2173 - Agro-biogeosystems: controls, feedbacks and impact
(POF4-217)},
pid = {G:(DE-HGF)POF4-2173},
typ = {PUB:(DE-HGF)16},
UT = {WOS:001480466600003},
doi = {10.1109/JSTARS.2025.3557956},
url = {https://juser.fz-juelich.de/record/1041302},
}