Figure 11. Comparison between the classification maps of the different models in dataset 2.
Figure 12. Comparison between the classification maps of differentscene sizes in dataset 2.
Figure 13. Comparison of the Kappa values and mutual information entropy between
different scene sizes in dataset 2.
entropy sharply increases from 60 m to 70 m. Therefore, in this
study, we selected 60 m as an appropriate scene size to balance
the detection accuracy and mapping detail.
Conclusions
In this paper, we have proposed a scene-based framework to
effectively detect tea gardens from high-resolution remotely
sensed imagery. The proposed framework is made up of three
scene classification models: the bag-of-visual-words model,
supervised latent Dirichlet allocation, and the unsupervised
convolutional neural network. These models achieved a high
accuracy and produced fine classification maps in our study
area. Through discussion and comparison, it was found that
unsupervised feature learning based on the
UCNN
outperformed
the other models in both accuracy assessment and visual
results. Furthermore, it was found that the supplement of the
textural features could significantly
improve the performance of the
topic models, but it had no effect on
the performance of the
UCNN
, since
this model can automatically and
adequately extract the discrimina-
tive features from the raw images.
To the best of our knowledge,
this is the first study of tea garden
detection using remotely sensed
imagery. Since tea plants are the
main cash crop in Chinese rural
areas, particularly in the mountain
villages in the southern provinces,
tea garden detection and monitor-
ing is important for the assessment
and guidance of economic develop-
ment in these areas. In the future,
we plan to implement the proposed
tea garden detection framework at
a larger scale for tea cultivation and
economic assessment.
Acknowledgments
The research was supported by
the National Natural Science
Foundation of China under Grants
41522110 and 41771360, the
Hubei Provincial Natural Science
Foundation of China under Grant
2017CFA029, and the National
Key Research and Development
Program of China under Grant
2016YFB0501403.
Figure 10. Comparison between the Kappa values of the
different models.
730
November 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
