Geography Reference
In-Depth Information
patterns. The NDVI value of the deciduous broadleaved forest is the highest in
those four types of vegetation, and that of open shrublands is the lowest. The
statistical curve from the classified land cover maps shows that the evergreen land
cover had no remarkable change during the study period. However, the deciduous
forest has a single peak in the sliding curve of NDVI in a year (Fig.
3.15
). The
possible reason is that the deciduous broadleaved forests are mainly located in the
temperate zone, while the needleleaved forests are mainly in a cold-temperate
zone or on mountains in a temperate zone.
3.3.4 Concluding Remarks on the Reclassified LUCC
In this chapter, this study aims to improve vegetation classification accuracy of the
land cover in North China by employing the technique of data mining different
satellite-derived land cover data of China, higher-precision land use data and other
ancillary spatial data. By computing the gain value of attributes for the vegetation
classification, the results showed that special monthly NDVI information is the
most important, and temperature is more sensitive than precipitation to the local
land cover changes. Vegetation classification method identifies the classes
including closed forest, shrubland and grassland with their exclusive spectral
feature parameters.
The accuracy of land cover classification is assessed by comparing the classi-
fication results with some reference data that is proved with actual land cover. In
this study, we find the accuracy of the C4.5 classifier is 88.96 %, which is higher
than others, including NBTree, SimpleCart, REPTree and BFTree. Besides, we
calculate the confusion matrix and ROC value of vegetation classification. The
kappa factor is 0.87 and the ROC value almost reach 0.90 in a lump sum, but the
ROC value of the deciduous broadleaved forest is only 0.74. The validation all
over China show that the overall accuracy of the land cover map is 83.14 %, which
is over 80 % and higher than that of other land cover maps matching requirement
of climate simulation work. Therefore, the results indicate provable improvement
of modeling accuracy for simulation of the land surface processes over China and
which can be introduced as parameters of dynamical downscaling into other
regional climate simulation.
To summarize, the developed classifier in this study can rapidly convert the
high resolution CAS land use types into the land cover types for climate simulation
work at regional climate models. Moreover, time-series NDVI and NPP data
retrieved from the remote sensing data can fast generate high resolution time-
series vegetation data and automatically recognize dynamic parameters input of
the regional climate model, which can also efficiently improve the accuracy of
regional climate simulation model. In addition, the results may provide supports to
other land surface scientific researches.
