dc.description.abstract | Rice is one of the most important crops in Taiwan. The government needs to know the planting status of rice every now and then, such as the planting location and planting area, for yield statistics and decision making. Traditionally, each remote sensing image is interpreted and digitized by manual annotation. In recent years, with the development of artificial intelligence-related technology, if we can use related technology to assist experts in interpreting remote sensing images, we can reduce the demand for human resources, reduce the possible misjudgment caused by manual interpretation, and improve the operational efficiency. Therefore, our team uses deep learning technology to generate a rice interpretation model that uses aerial images as input and outputs a classification result showing where are rice and non-rice., which will reduce the need for human resources to view and annotate rice on aerial images.
In the past, after obtaining the parcel vector map data, our team added the parcel vector map to the dataset, extracted the parcel information from the parcel vector map, based on the research result pixel-based UNet-VGG16 [1], tried to implement the parcel-based interpretation with minimal changes. Instead of modifying the UNet-VGG16 model, the design idea is to take out the feature map of UNet-VGG16 in the back-end network layer and processes it and parcel information to generate parcel-based data as input to another model FNN (Fully Neural Network) to achieve parcel -based interpretation, which has more accurate test results than the pixel-based UNet-VGG16. However, this way of using parcel information is too time-consuming in model training and testing, and is limited by the fact that the design of FNN requires a large amount of data that conforms to certain rules.
This study uses parcel information in a different way than the UNet-FNN, and uses parcel information directly to retrieve parcel -based image data from aerial image, and proposes a different network architecture from UNet-FNN, that is VGG16BN-G, to do the parcel-based interpretation on the parcel-based image data.
For the contribution of this study, the results of the same test design and t-test show that VGG16BN-G requires only about 20% of the training time and about 6% of the training data of UNet-FNN to achieve similar performance as UNet-VGG and UNet-FNNN without significant differences. The box plot shows that VGG16BN-G has similar stability to UNet-FNN and better stability than UNet-VGG16. Finally, based on the experience of our research team, we propose a guideline for aerial image preparation for rice interpretation, including an idea process to check the image quality. | en_US |