A deep learning based estimate of aboveground forest carbon density in
northeast China
Abstract
Quantifying forest biomass is an important part of determining the
regional carbon balance, but currently there is little knowledge
regarding forest carbon storage at a high spatial resolution. Here, we
combined a deep learning (DL) algorithm with field measurements, light
detection and ranging (LiDAR) observations, Landsat and ALOS/PALSAR
images to develop a spatially explicit estimate of forest aboveground
carbon density at a 30 m spatial resolution for northeast China, home to
nearly one-third of China’s forested area. We also conducted an
uncertainty analysis using the bootstrap method. The DL method performed
well, with a high coefficient of determination (R2) of 0.84 and a
relatively low root mean squared error of 6.28 MgC ha-1, and is superior
to traditional machine learning methods such as random forest, support
vector machine and artificial neural network. The forest carbon storage
is estimated to be 2.43 ± 0.10 PgC, and increases along the latitude
gradient. Among climatic factors, the wettest month precipitation and
annual mean temperature stand out in explaining the spatial variation of
forest carbon density, with contributions reaching 15.8% and 10.8%,
respectively. A model-data comparison shows that current ecosystem
models generally capture the spatial pattern of forest carbon density
but underestimate the forest carbon storage by 22.2%, partially due to
the overestimation of high-temperature inhibition, highlighting the need
to re-parameterize such temperature effects in forest carbon
simulations.