|
One of the biggest challenges in High Resolution Remote-Sensing Interpretation based on deep learning technology is the
access of huge number of sample annotations. Currently, the sample annotation is carried out manually, which is a boring,
time-consuming and labor-consuming work. Remote-Sensing Imagery comes from a variety of sources, each with its own
independent features, so that there is insufficient quantity of sample annotations to cover most of these features. The Third
National Land Survey (TNLS) data is a national land survey with high-quality manual annotation. This paper discusses
the use of TNLS data instead of manual sample labeling for model training, and three hybrid methods of sample cleansing,
data augmentation and training in turn, to create a deep learning training method of water identification. The results shown
that the TNLS data with three hybrid methods was able to identify high-quality waters, the Frequency Weighted
Intersection over Union (FWIOU) was up to 91.8%, which was 5.9% higher than the conventional training methods. The
deep learning model used in this research was proposed to obtain the high accuracy and saved a lot of human resources,
which had a wide range of practicability.
|
