Comparisons of Multi Resolution Based AI Training Data and Algorithms Using Remote Sensing Focus on Landcover

Abstract

The purpose of this study was to construct artificial intelligence (AI) training datasets based on multi-resolution remote sensing and analyze the results through learning algorithms in an attempt to apply machine learning efficiently to (quasi) real-time changing landcover data. Multi-resolution datasets of landcover at 0.51- and 10-m resolution were constructed from aerial and satellite images obtained from the Sentinel-2 mission. Aerial image data (a total of 49,700 data sets) and satellite image data (300 data sets) were constructed to achieve 50,000 multi-resolution datasets. In addition, raw data were compiled as metadata in JavaScript Objection Notation format for use as reference material. To minimize data errors, a two-step verification process was performed consisting of data refinement and data annotation to improve the quality of the machine learning datasets. SegNet, U-Net, and DeeplabV3+ algorithms were applied to the datasets; the results showed accuracy levels of 71.5%, 77.8%, and 76.3% for aerial image datasets and 88.4%, 91.4%, and 85.8% for satellite image datasets, respectively. Of the landcover categories, the forest category had the highest accuracy. The landcover datasets for AI training constructed in this study provide a helpful reference in the field of landcover classification and change detection using AI. Specifically, the datasets for AI training are applicable to large-scale landcover studies, including those targeting the entirety of Korea.