Land cover segmentation of aerial imagery using SegNet

Land cover relates to the biophysical characteristics of the Earth’s surface, identifying vegetation, water, bare soil or artificial infrastructure. Land cover mapping is essential for planning and managing natural resource, and for understanding distribution of habitats. Land cover classification for land cover mapping has been developed in a variety of ways. Among them, there are many attempts to classification land cover using deep learning techniques such as Convolutional Neural Network(CNN). CNN has been developed in many models, and semantic segmentation techniques that combining segmentation are also being announced. Among the Semantic Segmentation models developed until recently, SegNet has high accuracy and learning efficiency. We analyzed the availability of SegNet in the Land Cover classification. The study area was conducted in parts of South Chungcheongnam-do in South Korea. For the learning of the model, 2,000 data were constructed with the same size using the aerial image, and the constructed data was divided into training and validation data by 8 to 2. To solve the problem of class imbalance, which causes problems such as overfitting due to the difference in area per class, the weight value of each class was calculated using medium frequency balancing method. In order to calculate the hyper parameter optimization, the batch size was changed from 1 to 5 and the iteration was changed from 0 to 100,000 times Our experiments show that an overall accuracy (OA) of up to 85%, which confirmed the positive possibility of the semantic segmentation technique in the study of land cover classification.