Quantifying sources of error in multitemporal multisensor lake mapping

Regional-to global-scale lake maps can now be produced using existing technology and freely available data and serve as powerful tools for a variety of lake-and water-related studies. The accuracy of these studies depends in part on the accuracy of the lake map that they use. Mapping lakes using remote sensing requires a careful study of error and uncertainty. Errors in lake maps are caused by sensor-specific, lake-specific, and processing-specific factors. These can be further broken down to spatial, spectral/radiometric, and temporal factors. In this study, we analyse and compare these factors using modern and historical Landsat images along with intensive ground surveys of lakes in northern Alaska. Percentage error in lake area relative to lake size decreases for larger and more circular lakes, making a minimum size threshold an effective error mitigation practice. Image resampling involved in image transformation significantly increased error in lake area and is easily avoided by performing co-registration in the vector domain. Spectral properties varied for individual lakes due to depth, suspended sediment, vegetation, and other in situ factors, necessitating a normalized water index and independently derived threshold values for each lake. For lake change detection studies, spatially degrading a finer resolution image to the resolution of the coarser image a common practice does not significantly affect the difference in observed lake area. Due to the large numbers of lakes, particularly in the climatologically sensitive Arctic region, small errors in individual lake areas can compound to significantly impact results on regional to global scales. This study is intended to inform future static and multitemporal lake remote-sensing studies by evaluating errors and uncertainties in lake area, as measured by remote sensing.