Leaf, flowering and fruiting phenology of plants plays an important role in ecosystem processes and plant-animal interactions, including ecosystem responses to climate change and extreme events such as El Niño. In diverse tropical forests, variation in phenology among and within tree species is not well-characterized, largely because ground-based visual observations are labor-intensive for large numbers of individuals per species. Near surface remote sensing techniques allow for collection of frequent images of the canopy from which phenology metrics can be derived. We used Unmanned Aerial Vehicles (UAVs) to monitor phenology in a 50-ha forest dynamics plot on Barro Colorado Island, Panama every two weeks for a year. We delineated tree crowns in the UAV images and linked them to ground-based forest inventory data to provide the species identity of over 2,000 crowns. We calculated both mean color indices (e.g. widely used green chromatic color) and texture indices for each crown to distinguish leaf-on, leaf-off and flower-on phenological states. We created a training data of visually-determined phenological states for X crowns of two species to train a probabilistic classifier that would predict phenological state from the mean color and texture data for each crown.
Results/Conclusions
Image-analysis algorithms based on mean color, which have been successfully applied to time-series of canopy images (“phenocam” data) of mainly temperate forests, performed poorly when applied to UAV-acquired images at our tropical forest site. In contrast, including temporal patterns of texture variation in the classification algorithm performed well with a large number of common canopy tree species. Our analyses revealed previously undocumented patterns in canopy phenology, including a surprisingly long (~6 month) leaf-off season for many individuals of Tabebuia rosea, multiple deciduous periods for Ceiba pentandra, and large intraspecific variation for most species examined.