2.8. Performance of the classification algorithm
The classification accuracy was assessed using pixel-wise and
polygon-wise overall accuracy (OA) that were calculated for each
PT. The pixel-wise OA is the number of correctly classified pixels divided by the number of classified pixels. Polygon-wise OA is
the number of correctly classified polygons divided by the number
of classified polygons. A polygon was considered correctly classified if a majority of the pixels within the polygon was classified in
the same class as that polygons class. A polygon was considered
classified if a majority of its pixels were classified into some of the
target classes. Non-vegetation pixels were excluded from the calculations. The percentage of unclassified vegetation pixels in the
image was calculated by dividing the number of unclassified pixels in the image by the total number of pixels in image, while the number of non-vegetation pixels was excluded. The percentage of
unclassified pixels within target and non-target polygons was calculated similarly, with the additional exclusion of shadow polygons
and pixels. Additionally, the percentage of unclassified target and
non-target polygons were calculated. Omission and commission
errors were studied for selected classifications.
3. Results
3.1. Reflectance data and MNF transformed data
MNF transformation condensed spectral variation into smaller
set of bands (Fig. 2). The classification tests with 8–14 MNF bands
yielded the highest pixel-wise OA (90.8%) with the first 12 MNF
bands (Table 2). As a comparison, the pixel-wise OA achieved with
all 64 spectral bands was 71.1%. The first 12 MNF bands were thus
used in the later classifications.
3.2. Effect of probability threshold
The pixel-wise OA remained constant (90.8%) for up to PT
0.3 while all pixels were classified (Fig. 3a). As PT was further
increased, OA started to increase while the percentage of unclassified vegetation pixels in image increased, respectively. By visual
interpretation, it was noted that the validation pixels that were first
left unclassified were on the edges of tree crowns and in other areas
where it was hard to find pure pixels, i.e., pixels which represent
only one species without shadowing or major soil effects. Polygonwise OA increased slightly slower than the pixel-wise OA as the
PT was increased. The percentage of unclassified pixels increased
faster within non-target polygons than within target polygons as
PT was increased (Fig. 3b). Similarly, the percentage of unclassified non-target polygons increased faster than the percentage of
unclassified target polygons (Fig. 3b).
3.3. Classification maps of the study area
The classifications reveal the general distribution patterns of the
studied crops and tree species in the study area (Fig. 4a and b).
Table 2
The pixel-wise overall accuracy (OA) for classifications with the first 8–14 minimum
noise fraction (MNF) bands.
Number of bands 8 9 10 11 12 13 14
OA (%) 89.8 90.0 90.3 90.4 90.8 90.3 90.0
Table 3
Pixel-wise confusion matrix with probability threshold (PT) of 0.0.
Validation data
Classification Aca Ban Gre Mai Man Sha Sug Yam Total
Unclassified 00000000 0 Acacia 325 0 29 3 10 1 0 12 380
Banana 2 312 0 1 19 1 23 7 365