Seminar: Wenli Zhao

Evaporative fraction (EF), which is defined as the fraction of available energy portioned toward the latent heat flux, has a strong link with soil moisture availability and could be the limiting factor of latent heat flux. In comparison to non-temporal models, our machine learning model, which incorporates a memory effect, emphasizes the significance of memory effect in predicting EF. Utilizing our model as a foundation, we employ the Integrated Gradients (IG) score – a widely-used explainable machine learning method – to unpack the data-driven model and explore the temporal drivers of EF dynamics during soil moisture dry-down events. Our analysis of two sites (one grassland and one woody savanna) with longer-term observations reveals preliminary findings. It suggests that grasslands, characterized by shallow rooting depths, tend to rely more on the most recent time-series data in contrast to woody savannas with relatively deeper rooting depths. Furthermore, an intriguing finding is observed for the days immediately following precipitation events, where the impact of radiation and precipitation seem equally significant. This may be attributed to the fact that during this period, when soil moisture exceeds field capacity, energy becomes the primary limiting factor. Conversely, as soil moisture content drops below a critical threshold, the importance of recent rainfall becomes the main drivers, as moisture becomes the principal limiting factor. This preliminary analysis suggests the potential to identify not only the overall spatial drivers of EF but also the likely drivers of spatio-temporal variability. However, it is important to note that further analysis is required to validate the credibility of our current findings. We welcome any suggestions and insights in this regard.