Research on Adaptive Segmentation of Typical Objects in Remote Sensing Images

The high-resolution remote sensing image contains rich information of surface features. However, the same surface features collected in different regions at different times still have large appearance differences. This problem makes the network parameters trained in a single environment difficult to adapt to the extraction of similar objects in different environments. In order to solve these problems, this paper proposes a domain adaptive method based on CNN to solve the single class target segmentation problem in different feature domains for remote sensing images. The model consists of two parts. Firstly, this paper selects a segmentation framework suitable for remote sensing image feature extraction as a generator. Different feature layers of the generator are extracted to describe the source domain and the target domain image in multiple layers. Second, the features are used as the input of the discriminator to adversarial learning to distinguish whether the input is from the source image or the target image. In order to verify the effectiveness of the algorithm, this paper takes the extraction of planting greenhouses in remote sensing images as an example, first this paper proposes a data set for the extraction of planting greenhouses and carries out pixel level annotation. Then, the remote sensing image of the greenhouse in Chenggong District, Kunming City, Yunnan Province, South China was made into the labeled source data, and the remote sensing image of the greenhouse in the Inner Mongolia Autonomous Region of the North was used as the target data of the unmarked. The experimental results on the database show that the recall rate and accuracy of the proposed algorithm are good, and the MAE is 0.544, which is better than the existing feature extraction method based on migration learning.