Artificial neural network models applied to olive tree phenology in Italy reveal daily insolation control of budbreak

Olive (Olea europaea L.) trees are traditionally cultivated in the Mediterranean basin, providing both healthy food and ecosystem services, such as climate change mitigation and soil erosion control, particularly in arid areas. Despite its importance, olive phenology, as impacted by climate change, is under-studied. To tackle this gap, we assessed the potential of feed-forward artificial neural network models to predict five main olive phenophases (apex budbreak, inflorescence, flowering, pit hardening and olive maturation index 1) at their onset for cultivars ‘Picholine’, ‘Carolea’ and ‘Coratina'. The dataset was collected from seven sites across Italy during the years 1997-2000.  Due to gaps in the dataset, the models were initialized by supervised training with the subset of full phenological observations, followed by semisupervised training based on the full dataset and iterative estimations of the missing observations. The softmax activation function was used in the output layer by interpreting the incremental phenological transitions as proportional to probabilities. The networks with at least four hidden layers activated by the sigmoid function and trained with the momentum method and linearly-decreasing parameters were best performing (validation RMSE of 15.5 d and 17.1 d for ‘Picholine’ and ‘Carolea’, respectively). Daily insolation consistently improved budbreak prediction with respect to daily mean temperature, suggesting the operation of photoreceptor activation mechanisms. Inflorescence was better predicted when daily minimum temperature was added, consistent with a chilling-warm requirement mechanism. Flowering was less consistent, but mean temperature was a primary controlling cue. Therefore, each phenophase is likely controlled by different climate cues. When tested on two independent flowering dates in 2017 and 2018 from one of the sites , the best performing models for each cultivar gave median errors of 4.3 d, 12.1 d, 7.4 d and 3.7 d for the ‘Picholine’, ‘Carolea’, ‘Coratina’ and the combinaed ‘Picholine+Carolea+Coratina’, respectively. The worse predictions for 'Carolea' is likely due to the hypothesized sensitivity of this cultivar to climate change, that occurred in the years between the training and the testing observations. Therefore, the olive sensitivity to climate change could be strongly cultivar-dependent, which calls for more in-depth investigation in the future. The calibrated models can be used both as operational and hypothesis-testing tools to study climate change effects on olive phenology.