the period between 1996 and 2000 (Figure 8e). The enlarged
landscapes shown in Figure 8f highlight the visible textural
differences between deciduous rubber plantations and natural
evergreen forests.
Then the resultant maps of deciduous rubber plantations
were validated using a confusion matrix based on the
ROIs
(Table 1). The overall accuracies of maps for 2016, the early
2010s, late 2000s, early 2000s, and late 1990s were 94%,
90%, 95%, 94%, and 92%, respectively, while kappa coef-
ficients were 0.93, 0.88, 0.93, 0.96, and 0.89. Similarly, user's
accuracies for deciduous rubber plantations were higher than
98% in 2016 and within the four epochs, while producer's ac-
curacies were 92%, 85%, 90%, 94%, and 85%, respectively.
These results suggest that the tri-temporal
CRNBR
method is
effective in distinguishing deciduous rubber plantations from
natural evergreen forests.
Figure 7. Maps of deciduous rubber plantations for the five time periods between 1991 and 2015: (a) early 1990s, (b) late
1990s, (c) early 2000s, (d) late 2000s, and (e) early 2010s. (f) The area change of deciduous rubber plantations during the five
different periods.
Figure 8. Spatial distributions of validation
ROIs
for (a) 2016, (b) 2011–2015, (c) 2006–2010, (d) 2001–2005, and (e) 1996–2000.
(f) Enlarged examples of
ROIs
for deciduous rubber plantations (ring shape) and natural evergreen forests with
GE
images.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
September 2019
693
