INFLUENCE EVALUATION OF MACHINE LEARNING ACCURACY ON RIVERINE LAND COVER DETECTION FOR FLOOD FLOW ANALYSIS

Abstract

 This paper examined the influence of machine learning detection accuracy of riverine land covers using satellite images on their application of flood flow simulation. A planar 2D flow model was employed for the accuracy examination with riverbed roughness detected from the riverine land cover classification by random forests. Examined here were two cases with different machine learning training for a 2 km channel section in the Kurobe River. The result indicated that the flood flow model in this study could simulate both the velocity field and water surface profile with an accuracy high enough for practical flood flow analysis when the F-measure of the riverine land cover detection exceeded 0.8. It could suggest the practical applicability of F-measure for riverine land covers as an accuracy index necessary for the flood flow simulation.