Accurate Land Cover Classification via Training and Validation Data Refinement Using DNRS

Yoshie Ishii*, Junichi Susaki, Tsuguki Kinoshita, Koki Iwao

Abstract

Land cover maps provide critical insights for a variety of applications, including monitoring natural disasters and hazards, assessing climate change impacts, and forecasting future environmental conditions. The accuracy of these maps significantly depends on the quality of the training data used in their generation. In addition, generating validation data is essential for creating land cover classification maps to demonstrate the accuracy of the maps. However, acquiring high-quality training and validation data is time and labor intensive as well as presents potential for error in data collection, interpretation, and ground surveys. This study introduces a novel training and validation data refinement method employing the directional neighborhood rough set approach to address these challenges. We applied this refinement method to training and validation data for land cover classification using Landsat-8 and FLC1 datasets. The results demonstrate that the proposed method effectively identifies reliable training and validation data, thereby enhancing the quality of land cover maps and providing the assessment method with several confidence levels depending on the purposes.

Keywords: Class Boundary, Directional Neighborhood Rough Set, Land Cover Classification, Training Data, Validation Data

Full Text: