Traditional remote sensing methods are rarely applied to headwater streams because of a mismatch in spatial scale. With improved technology, remote sensing using unoccupied aerial vehicles enables imagery to be collected at high spatial and temporal resolution. However, these innovations have been used less frequently to detect changes within the water column. In addition, it is unclear whether classification methods developed along a single reach can be transferred through space (to other reaches) and through time, or how much information is needed to perform such classifications. This study combines methods of remote sensing, image processing, and machine learning to classify land cover and submerged aquatic vegetation along four urban stream reaches (170-325 m). A linear discriminant analysis (LDA) model was developed to provide land and water cover classification maps using training data.This method proved to be robust when classifying land cover along a single reach with minimal training data required. Using 200 pixels per land cover class resulted in 100% classification confidence for more than 93% of the pixels; when reducing this to only 50 pixels per land cover class, LDA yielded 100% classification confidence for nearly 80% of the total pixels. Through attempts to transfer relationships across reaches and flight dates, we found a greater percentage of misclassified pixels (upwards of 45% misclassification for flights on one date) when classifying across reaches on a single date as opposed to classification along a single reach for multiple dates. Overall, we recommend passive optical approaches, like the one used here, consider lighting conditions, reach orientation, and shading in data acquisition planning to mitigate the uncertainty they may introduce in delineating water column quality and cover.

No abstract