An example of how data quality hinders progress: translating the latest findings on the regulation of leaf senescence timing in trees into the DP3 model (v1.0)

Abstract. The timing of leaf senescence ends the growing season of deciduous trees, affecting the amount of atmospheric CO₂ sequestered by forests. Some climate models integrate the timing of leaf senescence, which can be simulated with process-oriented models. Here, we developed a process-oriented model of leaf senescence (the 'DP3 model') by testing 34 formulations of the leaf development process. The period between leaf unfolding and leaf senescence was separated into three subsequent phases with particular reactions to aging and stress, (sum of cold, photoperiod, and dry stress). The DP3 model and the compared previous models were equally accurate, but less accurate than the Null model (i.e., constant simulation of the mean observation of the calibration sample). This lower accuracy was very likely due to noise in the visually observed leaf senescence data, which blurred the signal of the process of leaf senescence, and incorrect model formulations. The leaf senescence data were attributed to most of the variation in the model error of the models compared, which was similarly affected by climatic and spatial deviations from the calibration sample across models. The DP3 model considerably contrasts previous models, allowing the development of new hypotheses, e.g. on the cause of senescence induction. Independently from model formulation, noisy leaf senescence data likely force the models to resort to the mean observation, impeding inferences from accuracy-based model comparisons about the process of leaf senescence. This implies the usage of data from as few sites as possible to minimize the noise due to different observers and small sample sizes when evaluating and further developing models of leaf senescence. Moreover, revised observation protocols should explain how to measure rather than to estimate the timing leaf senescence, e.g., based on greenness, involving digital cameras and automated image assessment.