Automated River Width Extraction from Remotely Sensed Images

In recent years many methods have been developed and applied for identifying water bodies, and mapping surface water using high-resolution and heterogeneous remote sensing images. In traditional indexing methods, multitemporal images provide spectral information that directly relates to water occurrence, but they have a relatively low resolution to obtain accurate physical information of river systems. Natural true-color images are available at relatively higher resolution, but they are not viable for detecting water boundaries primarily due to the discrepancies in illumination. For this reason, many models need user-defined ancillary data, complex rule-based expert systems, or new types of deep learning algorithms. The primary focus of this study is to develop a fully automated tool for identification of river wetted area from remotely sensed data. The method combines multitemporal and true-color image features. First, region-based Landsat images are synthetically enhanced and the water areas are identified using spectral water indexing. Then the water distance map is obtained by using squared Euclidean distance algorithm. Finally, a water mask is applied on a region-based true-color image of the river reach to distinguish water regions from no-water regions using pixel-based RGB information. The river centerlines using the Laplacian filter, and the river widths using the Euclidean distance algorithm are obtained without ancillary data. The integration of a convolutional neural network (CNN) algorithm into the river area identification tool is also underway. The performance of the method is tested using measured river surface data along a reach of the Tallahatchie River near Money, MS, which will also be used for training the CNN algorithm.