Histogram-based edge detection as a tool for detecting river coastline

Flood management is very important task in the context of rapid climate changes. Increasing the frequency of extreme weather and fluvial phenomena, such as droughts, water shortages or floods determines that detecting water bodies and boundaries between them and surrounding surface is an important and challenging issue. We elaborated a new approach for delineating river coastline based only on close-range RGB nadir images acquired by means of UAV (unmanned aerial vehicle), converted to HSV (hue, saturation, value) color space. We used spectral characteristics of water surface which has uniform V component, while another land cover types have heterogeneous V. Areas, where character of V changes considerably, are suspected to be  river coastline. Every aerial image was divided into 250 x 250 px cells, within which we calculated some statistical values (kurtosis, concentration around mode) in order to characterize shape of empirical distribution. Among others, we focused on identifying multi-modal or leptokurtic histograms. Results show that the detection rate (also known as the producer accuracy) ranges from 22,22% to 92,00%, while the false hit rate (also known as error of commission) ranges from 5,00% to 82,76%. For 70% of all analyzed images, presenting both narrow (10 m) and wide (more than 100 m) rivers, the detection rate was above 50%. Considering the subset of photos presenting only wide rivers, detection rate above 50% occurred for 75% of these images. For these cases, 56% of images do not exceed the false hit rate above 40%. The research is supported by the National Science Centre, Poland, through the project no. 2020/38/E/ST10/00295.