Author(s): Shijun Pan; Keisuke Yoshida; Takashi Kojima
Linked Author(s): Keisuke Yoshida
Keywords: No keywords
Abstract: In recent years, due to the increase of flooding caused by global climate change and the rapid development of artificial intelligence (AI) technology, hydraulic researchers have started to try to cope with the land cover classification (LCC) problem aided by deep learning (DL) techniques. To clarify the vegetation distribution in Asahi River, Japan, this study firstly utilized a DL-based DeepLabV3+ model for semantic segmentation of aerial plane-derived orthophoto. Although this orthophoto-based (RGB-based) method has advantages in producing LCC mapping, how to create reliable true label (TL) mapping efficiently is still a complex issue that this method is confronting with. Especially, in the process of creating TL maps, facing certain parts where fieldwork by labors is not possible, like hazardous field observation environments. And at the same time, maybe aerial photographs are not enough to extract the feature of the land cover, like grass and bamboo. Based on this mentioned situation, airborne laser bathymetry (ALB) datasets including voxel-based laser points (n) and digital surface model data minus digital terrain model data (l), can be reference of creating TL maps. Except of the TL map reference, we also modified the existing DeepLabV3+ model by connecting data at its result output port to access ALB datasets, including n and l, is named as RGBnl-based method. This paper mainly concentrates on how to classify riverine land cover using orthophoto with the aid of the ALB dataset. In this study, an orthophoto of the corresponding area needs to be selected in advance, an orthophoto is synthesized by the corresponding coordinates (middle of the first pixel in the upper left corner), and finally the redundant part of this aerial photograph is whitened. As an experience of distributing DL dataset, this orthophoto needs to be divided into three parts (pixel ratio of 8:1: 1): train-, valid- and test-part. The results show that the RGB-based and RGBnl-based methods are 0.89,0. 84 and 0.88,0. 84 in terms of overall accuracy and macro F1 scores derived from test-part confusion matrix, respectively. Therefore, this also concludes that our TL map is more plausible for both methods.
Year: 2022