Improving the accuracy of tidal flat vegetation classification using UAV and deep learning

Abstract

The salt marsh system on the tidal flat is one of the most productively ecological wetlands with high biological productivity and blue carbon sequestration levels. In recent, the advancement of deep learning (DL) technology and Unmanned Aerial Vehicles (UAVs) has made it feasible to monitor tidal flat vegetation more efficiently and precisely. However, many studies have been conducted at the landscape level, and little is known about the performance of species discrimination in very small patches and in mixed vegetation. In this Study, we constructed a dataset based on UAV-RGB data and compared the classification performance of traditional Pixel-Based (PB) and DL methods. The Classification accuracy were compared with the five scenarios (combination of labeling data size and annotation type). As expected, the results showed that in comparison with the PB and DL methods, the DL method achieved the most accurate classification results. However, there was no significant difference in OA between the two annotation types and labeling data sizes in DL methods. Nevertheless, it was confirmed that the polygon type annotation method was more effective in the mixed vegetation classification than the bounding box type. Moreover, it was also confirmed that the smaller size of labeling data was more effective for detecting the small vegetation patches. The results suggested that a combination of UAV-RGB data and DL can facilitate long-term and accurate monitoring of coastal wetland vegetation at the local scale. Furthermore, accurate classification of species was possible, and this information is expected to be used for more accurate blue carbon estimation studies in the future.