8
Q .W a n ge ta l .
Airplane
Beach
Agricultural
Baseball diamond
Buildings
Chaparral
Denseresidential
Forest
Freeway
Golfcourse
Intersection
Mediumresidential
Mobilehomepark
Overpass
Parkinglot
Tenniscourt
River
Runway
Sparseresidential
Storagetanks
Harbor
Airplane
Beach
Agricultural
Baseballdiamond
Buildings
Chaparral
Deneseresidential
Forest
Freeway
Golfcourse
Harbor
Intersection
Mediumresidential
Mobilehomepark
Overpass
Parkinglot
Tenniscourt
River
Runway
Sparseresidential
Storagetanks
Fig. 2. Confusion matrix for 21-category performed result of the proposed method
(86.78%)
that our proposed method enhances the classification effect 1.69% over the bestexisted results in [8,12–14]. To sum up, our approach achieves a more favorable
effect on this dataset because of the combination of feature fusion and PCA
transformation.
Table 3. Overall classification accuracies on dataset
Approaches
Overall accuracies (%)
State-of-art BOVW [12]
76.81
Texture [12]
76.91
spck++ [14]
77.38
Approach of [13] 75.33
Strategy 1 of [8] 85.09
Our method
86.78
4 Discussion of Time
In this part, the time spent on the training process with different classifiers
is compared. As Table 4 shown, applying Caffe framework to train the CNN
architectures, VGGNet-16 and ResNet-50, based on GPU for acceleration need
4376 and 2437 s, individually. However, our proposed method with random forest
for classifying based on the 21-category dataset just needs 24.81 seconds, which
shortens over one hundred times. In addition, the classification accuracies of our
method are 86.78%, which is higher than VGGNet-16 and ResNet-50. Moreover,
after analyzing, PCA transformation helps to reduce the time for training, too.