[ACG54-08] Data-Driven Estimation of Soil Respiration in Japan
Keywords:Soil Respiration, Carbon Cycle, Observation Network, Machine Learning, Remote Sensing
concentration and climate changes. Yet, spatial and temporal patterns of atmosphere-land CO2
exchanges were not sufficiently clarified. Among various processes of terrestrial carbon fluxes (e.g.
photosynthesis and respiration), fluxes relating to soil is one of the least understandable ones. Soil
Respiration (SR), the sum of root respiration and heterotrophic respiration, is one of the most
essential components of soil carbon cycles. So far, various efforts were conducted to understand
SRs. Many observation stations directly measure SR using chambers. Using these observation data
and literature survey, several studies estimated spatial and temporal patterns of SR at global scales
based on simple semi-empirical equations. However, the database (Soil Respiration Database;
SRDB) used in these global scale studies contains inconsistently observed datasets. These
inconsistency will produce additional uncertainties in estimated fluxes.
The largest SR observation network across Asia developed and maintained by NIES, Japan
can be a good candidate to estimate spatial and temporal variations in SR across Asia. These
observations are conducted with consistent observation protocol and quality controls. Therefore, we
attempted to estimate SR in Japan with observation data which are in one of the largest data set
unified their methods, and remote sensing data by using machine learning. We used eight sites data
across Japan and conducted empirical (data-driven) upscaling using random forest regression and
MODIS sensor data.
We confirmed that our approach reasonably estimates spatial and temporal variasions in
SR across Japan. We tested two experiments at site-level using (1) remote sensing based input data
only and (2) remote sensing based input data and site measurements (soil temperature and air
temperature). Site-level experiments shows good performance of the model (e.g. R= 0.73 – remote
sensing only and R=0.79 remote sensing + site observation) even if input data are remote sensing
based only. Spatial and temporal estimation also shows reasonable seasonal variation in SR.