Programs to monitor lake area change are becoming increasingly important in high latitude regions, and their development often requires evaluating tradeoffs among different approaches in terms of accuracy of measurement, consistency across multiple users over long time periods, and efficiency. We compared three supervised methods for lake classification from Landsat imagery (density slicing, classification trees, and feature extraction). The accuracy of lake area and number estimates was evaluated relative to high-resolution aerial photography acquired within two days of satellite overpasses. The shortwave infrared band 5 was better at separating surface water from nonwater when used alone than when combined with other spectral bands. The simplest of the three methods, density slicing, performed best overall. The classification tree method resulted in the most omission errors (approx. 2x), feature extraction resulted in the most commission errors (approx. 4x), and density slicing had the least directional bias (approx. half of the lakes with overestimated area and half of the lakes with underestimated area). Feature extraction was the least consistent across training sets (i.e., large standard error among different training sets). Density slicing was the best of the three at classifying small lakes as evidenced by its lower optimal minimum lake size criterion of 5850 m2 compared with the other methods (8550 m2). Contrary to conventional wisdom, the use of additional spectral bands and a more sophisticated method not only required additional processing effort but also had a cost in terms of the accuracy and consistency of lake classifications.
