On water surface delineation in rivers using Landsat-8, Sentinel-1 and Sentinel-2 data

摘要

This study is a pilot project for the "San Francisco Flood Plain Project" (SFFPP), meant to delimit flood plain areas owned by the Brazilian federal government. The objective is to determine the attainable accuracy in river water surface delineation using satellite imagery from Landsat, Sentinel-1 and -2. We prioritize the evaluation of Landsat data due to its long systematical time series, allowing hydrological analysis requiring observations of at least 40 to 60 years and data from the Sentinel missions with their high frequency of revisit, improved spatial resolution (compared with Landsat) and possibility of observation in wet season (S-l). In our approach, we evaluated the accuracy by spectral bands individually and in combination, as well as polarization. We also tested a number of thematic information extraction techniques unsupervised (K-means and EM Cluster Analysis) and supervisioned (Random Forest-RF, k-nearest neighbors-KNN, Maximum Likelihood Classification-ML, Support Vector Machine-SVM, Mahalanobis). To validate our results, we used a PlanetScope mosaic (3 m). Results indicate that shortwave infrared bands have a higher capacity to separate water surface from other classes. For SAR data, the best separation was obtained by VV polarization (compared with VH). Techniques all reached agreement values>94% for the Sentinel-2 image,>93% for the Sentinel-1 image and>86% for the Landsat-8. We consider both methodologies effectives to extract the water surface and appropriate for the real estate issues of the SFFPP project.