UAV-based classification of mixed salt marsh vegetation using U-Net

Abstract

Remote sensing is very useful for mapping and monitoring of salt marsh vegetation. In the recent years, the rapid development of Unmanned Aerial Vehicle (UAV) technology with its high portability and ultra-high spatial resolution, has filled the gap between traditional field surveys and satellite monitoring. Here, we propose using a U-Net model applied to UAV image, for an accurate classification of mixed salt marsh vegetation. We compared the performance of pixel-based classification and deep learning methods based on UAV-RGB data. This study takes the coastal salt marsh of southwestern part of the Korea. Phragmites communis (reed), Suaeda maritima widely distributed in the study area. S. maritima gradually turns red as the altitude of the topography increase. DJI Matrice 300 RTK with the Zenmuse P1 RGB sensor were used. Orthomosaic image were generated with a ground sampling distance (GSD) of 0.6 cm and 5 cm Digital Elevation Model (DEM). The RGB images were used to generate training samples by visual interpretation for Maximum likelihood classification (MLC) and U-Net model. From the MLC method, it was difficult to classify due to morphological and spectral similarities between reed and S. marritima. The reeds were often misclassified as greenish S. marritima, and sediments were misclassified as reddish S. marritima. On the other hand, the method using the U-Net model showed a significant decrease in misclassified pixels. Despite the increased accuracy, there was still a tendency that very small patches of S. marritima were classified as sediment and the edge of the reed patch was misclassified as S. marritima. Considering the training of deep learning models usually requires a large amount of labeled data, how to efficiently and automatically generate deep learning labels for remote sensing classification becomes an important and pending issue in future research