Identifying Leaf Phenology of Deciduous Broadleaf Forests from PhenoCam Images Using a Convolutional Neural Network Regression Method

Vegetation phenology plays a key role in influencing ecosystem processes and biosphere-atmosphere feedbacks. Digital cameras such as PhenoCam that monitor vegetation canopies near real-time provide continuous images record phenological environmental changes. There is need to develop methods for automated effective detection of dynamics from images. Here we developed method predict leaf deciduous broadleaf forests individual using deep learning approaches. We tested four convolutional neural network regression (CNNR) networks on their ability growing dates based at 56 sites North America. In the one-site experiment, predicted dated after leaf-out events agree well with observed data, coefficient determination (R2) nearly 0.999, root mean square error (RMSE) up 3.7 days, absolute (MAE) 2.1 days. The achieved lower accuracies all-site experiment than R2 was 0.843, RMSE 25.2 MAE 9.3 days experiment. model accuracy increased when used region interest rather entire inputs. Compared existing rely time series studying phenology, found feasible solution identify