Learning to predict channel stability using biogeomorphic features

Current human land use activities are altering many components of the river landscape, resulting in unstable channels. Instability may have serious negative consequences for water quality, aquatic and riparian habitat, and for river-related human infrastructure such as bridges and roads. Resource management agencies have developed rapid bioassessment surveys to help assess stability in a fast and cost-effective way. While this assessment can be done for a single site fairly rapidly, it is still time-consuming to apply over large watersheds and assessment activities must be prioritized. We constructed a system that employs commonly available map data as inputs to cost-sensitive variants of decision tree algorithms to predict the relative channel stability of different sites. In particular, we use bagged lazy option trees (LOTs) and bagged probability estimation trees (PETs) to identify all unstable channels while making the smallest number of errors of classifying stable channels as unstable, thereby minimizing cost and maximizing safety. We measured the performance of the classifiers using ROC curves and found that the PETs performed better than the LOTs in situations where the number of instances of the stable and unstable classes were relatively balanced, but the LOTs did better where unstable examples were relatively rare compared to stable, perhaps due to the LOTs' ability to focus on individual examples.