In order to analyze the difference between aerial photos and land use status maps, machine learning techniques for image classification related to the realization of a learning model for aerial photos in image format and learning materials for vector ...
In order to analyze the difference between aerial photos and land use status maps, machine learning techniques for image classification related to the realization of a learning model for aerial photos in image format and learning materials for vector format land use status maps and image format aerial photos Spatial analysis is required for construction. As a method of using CNN (Convolutional Neural Network), it is possible to consider a method of building a new model using learning data and a method of using an existing model based on CNN. In the meantime, a number of models such as VGG16, CaffeNet, and GoogLeNet that have extended CNN as methods related to image classification have been developed. These are publicly available in an open source form, and various studies have been conducted on image classification and object detection for transfer modeling using this method.
The research stage can be roughly divided into three stages: the stage of constructing a transfer learning model for images, the stage of applying the model to vector data, and the search for inconsistent regions. Aerial image data, which is the data of the study, can be collected by using the website or by taking direct pictures. In order to collect data using the website, aerial photos and orthographic images provided by the National Land Information Platform (http://map.ngii.go.kr) can be used, and as a direct shooting method, an orthographic image method using a drone is possible. do. The website data collection enables data collection at various scales over a wide area but has a characteristic that a specific shooting time is determined. On the other hand, in the case of drone images, it is possible to acquire data on the latest changes, but there are limitations in the regional scope of imaging. Two methodologies of class classification and object extraction are applied to construct transfer learning model. Among the transfer learning models based on CNN, the V3 inception model is applied to the one that has the most use in object classification. The V3 induction model, also known as GoogLeNet, developed at Google Research Lab and opened as an open source, has many examples of application to satellite image classification. YOLO is applied as an object classification method. YOLO, developed at Washington University's research institute, is a famous model for real-time object detection transmitted to CCTV, and has the most use in the field of object detection, and many studies have been conducted in the field of spatial object detection in satellite images.
In the study of machine learning, the most representative field of data science, it takes a lot of time to collect suitable learning data and search to increase the accuracy of the model.
In the case of image classification, in the case of images acquired by general photography, the work is simple by labeling the photos, whereas in the case of aerial photographic images, the processing of spatial data should be considered because it is a vertical image of the ground rather than the side and has a coordinate system.