A smart classifier for extracting environmental data from digital image time-series: Applications for PhenoCam data in a tidal salt marsh

PhenoCams are part of a national network of automated digital cameras used to assess vegetation phenology transitions. Effectively analyzing PhenoCam time-series involves eliminating scenes with poor solar illumination or high cover of non-target objects such as water. We created a smart classifier to process images from the “GCESapelo” PhenoCam, which photographs a regularly-flooded salt marsh. The smart classifier, written in R, assigns pixels to target (vegetation) and non-target (water, shadows, fog and clouds) classes, allowing automated identification of optimal scenes for evaluating phenology. When compared to hand-classified validation images, the smart classifier identified scenes with optimal vegetation cover with 96% accuracy and other object classes with accuracies ranging from 86 to 100%. Accuracy for estimating object percent cover ranged from 74 to 100%. Pixel-classification with the smart classifier outperformed previous approaches (i.e. indices based on average color content within ROIs) and reduced variance in phenology index time-series. It can be readily adapted for other applications.