Prior studies of leaf phenology in tropical forests, whether based on fixed RGB cameras or Unoccupied Aerial Vehicle (UAV) overflights, have been limited to seasonal and spatial changes in leaf amount but not age. Yet leaf age is an important driver of seasonality in Gross Primary Productivity in tropical forests. We address this limitation by leveraging high spatial resolution (2 cm) RGB-UAV orthorectified mosaics to classify and quantify fractional composition of hundreds of individual tree crowns across three Central Amazon upland evergreen forests with different soils and structures. Crown components were young leaves (light-green), mature/old leaves (dark-green) and bare branches. A Random Forest classifier trained on a single UAV mosaic was applied across different seasons and forest types. High classification accuracy required simple but essential quality controls: homogeneous illumination, slight underexposure, fixed color balance, fixed shutter speed, and intercalibrating all mosaics to the training mosaic by histogram matching using Cumulative Distribution Function. The best model proved versatile, achieving >90% accuracy on independent test data – regardless of season, forest type, or distance from training site. Monitoring of 453 crowns revealed pronounced seasonal dynamics: greater abundance of young leaves and briefly bare branches during the drier months. We also detected fewer young leaves in valley bottom than plateau forests and reduced dry season leaf flush four months after the end of a severe drought. These patterns reveal strong seasonality and semi-synchronized leaf flushing across Central Amazon canopy species, providing a robust framework to study tropical forest productivity and climate sensitivity. • First classification of tropical forest crown cover fraction by leaf age class. • Single training set was versatile across seasons and forest types: >90% accuracy. • Camera settings, homogeneous illumination and CDF intercalibration are critical. • Best classifier predictors were robust to pixel illumination intensity.
Simonetti et al. (Mon,) studied this question.