Predicting Vegetation Phenology using Machine Learning based on Wavelet Transform of Meteorological Drivers

Climate change and extreme weather events have far reaching consequences for terrestrial ecosystems, particularly for vegetation phenology. However, the effects of meteorological variations on phenology are still not well understood, rendering phenology modeling a major challenge. Here we adapt explainable machine learning (ML) techniques from computer vision to investigate the role of meteorological variability and its multi-scale memory on phenology.
Specifically, we develop a modelling framework using convolutional neural networks trained on wavelet transformed key meteorological variables to predict vegetation greenness. The wavelet transformation of the meteorological time series (temperature, soil moisture, and shortwave radiation) yields two-dimensional images that reflect their different frequencies across a broad spectrum from multi-year variability to synoptic time scales. We use the green and red chromatic coordinate (GCC and RCC) from the ground-based PhenoCam network as proxies for the daily state of vegetation phenology. Additionally, to compensate for calibration artifacts across the sites, we use the satellite-based normalized difference vegetation index (NDVI) for normalisation.
Explainable ML techniques, such as Integrated Gradients, in combination with the wavelet images give us insight into the importance of the various meteorological factors as well as the length and timing of the weather events for the prediction of phenology. We present first results of our modelling framework and illustrate the effects of meteorological variability, with an emphasis on spring phenology, at different time scales. In particular, we use the interpretability of our model architecture to develop hypotheses and test them with manipulation experiments. In addition, we explore the model's ability to spatially extrapolate to unseen locations during training.
Such studies are important to understand the impact of climate change on the seasonal cycle of terrestrial ecosystems and to find out whether ML with explainable techniques can lead to a better understanding and thus improvements in modelling phenology.