Geography Reference
In-Depth Information
study find that the C4.5 algorithm is suitable for converting land use data to land
cover data of IGBP classification. We reset the decision rules with Net Primary
Productivity (NPP) and Normalized Difference Vegetation Index (NDVI) as the
indicators. The dataset with the accuracy of 88.96 % is divided into 66 % training
data and 34 % testing data. The validated accuracy of the classified land cover data
is 83.14 % by comparing with the Terrestrial Ecosystem Monitoring Sites (TEMS)
and high resolution images. Therefore, we may produce the temporal land cover
data using this method, which can meet the accuracy requirement of climate
simulation and that can be the parameters of dynamical downscaling in regional
climate simulation.
Many land cover data of China have been produced in recent years with the
remote sensing data. The previous study showed that the result of the precipitation
study would be greatly influenced if the accuracy of land cover data is under 80 %,
and the result may be worse as the accuracy continue to decrease (Ge et al.
2007
).
Unfortunately, neither the overall nor class-specific accuracy of most datasets can
meet the common requirements of the regional climate modeling. Therefore, it is
necessary to produce the land cover dataset with high accuracy for the climate
simulation based the existing land use dataset, land cover datasets and some
ancillary datasets. These available data with a high level of uncertainty may be
improved by the combining the different data sources so as to meet the require-
ment of climate simulation.
The researches on the climate modeling vary substantially in the spatial and
temporal scales. So the temporal land cover datasets are essential to the develop-
ment of a cohesive climate model. The CAS has constructed a land use dataset that
includes the data of 1988, 1995, 2000 and 2005 (Liu et al.
2003
). However, there are
still no comparisons of land cover datasets at the regional scale, especially in China
where the land use is changing drastically due to rapid economic development and
anthropogenic disturbance. Many studies have indicated that the disagreement
among the land cover datasets primarily resulted from the differences of sensors,
spatial resolutions, algorithms, and classification schemes (Kaptué Tchuenté et al.
2011
); Among them, the difference in the classification schemes is considered to be
the key reason for disagreement of the land cover datasets and the main obstacle to
comparing different land cover datasets. Therefore, it can make great contribution to
climate change research if we can take advantage of the long-term land use datasets
from the CAS, and use an appropriate method to convert them to the International
Geosphere Biosphere Programme (IGBP) land classification scheme. It consists of
seventeen categories (Table
3.7
) and is widely accepted and used in the simulation
of climate changes (Gao and Jia
2012
).
The decision tree is one of the most powerful classification algorithms to
classify land cover type of remote sensing image (Simard et al.
2010
). The
decision tree technique is more suitable for the analysis of categorical outcomes.
Besides, it is easy to interpret, computationally inexpensive and capable of dealing
with noisy data. Furthermore, its prediction model is more understandable to the
users. The decision tree classifiers include the C4.5/C5.0/J48, NBTree, Simple-
Cart, REPTree, BFTree, etc., among which the C4.5/C5.0/J48 classifier is the most
