[A-CG45] Water and sediment dynamics from land to coastal zones

[ACG45-P03] A 60cm water body map using aerial photography for Japan

*Megumi Watanabe¹, Dai Yamazaki¹ (1.Institute of Industrial Science, The University of Tokyo, Japan)

Keywords:Waterbody, High spatial resolution, RGB, Aerial photography, Bayesian inference, Hydrography

The river width is an essential parameter for many applications, such as flood inundation modeling. However, most current global river models calculate the river width of small rivers using empirical equations based on discharge and drainage area. Therefore, high spatial resolution river width data is required. Some of the latest optical satellites images, such as Planet and aerial photography, have attracted considerable attention for their detecting ability with high spatial resolution. Remotely observed data with these images can develop a high spatial resolution water body map. Approximately 60 cm resolution aerial photography data is available over entire Japan. However, aerial photography only contains the RGB portion of the visible light spectrum. Conventional methods to detect water bodies using remote sensing data require the infrared part in addition to the visible light spectrum. Here, we attempt to develop a 60cm water body map using aerial photography without the infrared radiation data for Japan. Based on the developed water body map, our final goal is to estimate river width with a high spatial resolution.
We employed a Bayesian inference to estimate water probability using the 60 cm resolution aerial photography and approximately 10 m to 30 m resolution water fraction by Landsat-based dataset “G1WBM” and geospatial dataset as prior probabilities of water. Also, the hydrography dataset “J-FlwDir” was combined as topographical information to separate objects with the same color as the water surface from the water body. Firstly, RGB information from the aerial photography was transformed into Hue-Saturation-Value (HSV) color-space. Secondly, each pixel of aerial photography was classified mainly based on this HSV color-space and used as a condition of the Bayesian inference. Consequently, the probability of water was calculated for each classified pixel. Our method enables us to detect small rivers not seen by the previous Landsat-based 30m water body map. Rivers located in open space and not covered by vegetation or other objects can be detected well. However, there remain false alarms mainly from vegetation. Further improvement of this issue is necessary. Ultimately, river width can be calculated using the developed 60 cm waterbody map and flow direction map.