geocomputational workflows for analysing spring plant phenology in space and time
Hamed Mehdi Poor is a PhD student in the Geo-information Processing (GIP) Department. His supervisor is prof.dr. R. Zurita-Milla from the Faculty of Geo-Information Science and Earth Observation.
Among the various research questions raised by climate change, the question: “how does climate change affect vegetation seasonality?” is crucial. This is because changes in vegetation seasonality have both global and substantial implications for our planet. In particular, phenology is the science that studies seasonal plant and animal life cycle events (phenophases) and how annual and inter-annual variations in weather and environmental conditions affect them. In this respect, volunteered phenological observations (VPOs) and phenological model outputs are key information to study phenology in space and time. This phenological information helps to better understand the impact of climate change on vegetation seasonality. In this research, we designed novel geocomputational workflows to explore spring plant phenology (SPP) at large scale and over long periods using VPOs and temperature-driven phenological models. The workflows focus on 1) checking the consistency of VPOs using contextual geo-information and domain knowledge and 2) analyzing the impact of the type of phenological model as well as of its input data sources and their spatial resolution on the patterns and trends derived from the model.
The first workflow uses the statistical distribution of the VPOs as well as environmental contextual information to provide robust consistency checks. It relies on all the available environmental information for the VPO sites (in our case weather variables) and it is suited for areas with strong environmental gradients. The workflow was used to identify inconsistent (i.e., anomalously early or late in relation to their associated environmental conditions) VPOs from the USA National Phenology Network (USA-NPN). In particular, it was applied to flowering observations of common and cloned lilac plants (Syringa vulgaris and Syringa x chinensis) in the coterminous United States for the period 1980 to 2013. About 97% of the VPOs were flagged as consistent, indicating that volunteers provided reliable information. Relative to the original dataset, the exclusion of inconsistent observations changed the apparent rate of change in lilac bloom dates by two days per decade, indicating the importance of inconsistency checking as a key step to analyze volunteered geographic information.
The second workflow uses variables that are related to the phenology domain. In this case, temperature as it is an important phenological driver in temperate climates and it strongly influences phenological synchrony. This knowledge forms the basis for defining, searching and optimizing consistency constraints. This workflow was tested using VPOs collected in the Netherlands during the period 2003–2015. The average percentage of inconsistent observations was low to moderate (ranging from 1% for wood anemone (Anemone nemorosa L.) and pedunculate oak (Quercus robur L.) to 15% for cow parsley species (Anthriscus sylvestris (L.) Hoffm)). This again indicates that volunteers provide reliable phenological information. We also found a significant correlation between the standard deviation of DOY of the observed events and the accumulation of daily temperature (with correlation coefficients ranging from 0.78 for lesser celandine (Ficaria verna Huds), and 0.60 for pedunculate oak). This confirms that colder days in late winter and early spring lead to synchronous flowering and leafing onsets. Our results highlight the potential of combining environmental and domain information and knowledge to check the consistency of (phenological) volunteered geographic information.
The third workflow was designed to analyse and compare patterns and trends from a suite of temperature-based phenological models, namely the Extended Spring Indices (SI‐x), Thermal Time and Photothermal Time models. These models were first calibrated using historical lilac leaf onset observations for the period 1961-1994. Then, contemporary VPOs and daily gridded temperature data were used to validate the models. Results show that the root-mean-square errors (RMSEs) of the SI‐x and Thermal Time models are similar, and about two days lower than those provided by the other models. Yet, the dates of leaf out provided by each of the models are up to 11 days different, and the trends are up to a week per decade different. These results also show that the statistical significance of phenological trends strongly depends on the type of model. Therefore, current approaches for validating phenological models based on global statistics such as RMSE do not provide information about the variability of patterns and trends in different regions. Studies using phenological models and gridded input data to study climate change impact on plant seasonality should check both the spatial and temporal variability of the model outputs at large-scale. Using a model that is found less valid than another one (i.e., with a “worse” RMSE for a given set of validation observations) may still provide more realistic patterns and trends when compared with large-scale phenological data and/or information.
The fourth workflow was used to evaluate how the source and spatial resolution (i.e., scale) of the input data might affect phenological models and indices that track variations and shifts of vegetation seasonality at continental scales. The workflow, based on cloud computing and volunteered phenological observations, focuses on the SI‐x, which estimate the day of year (DOY) for first leaf (FL) and first bloom (FB) in plants sensitive to accumulation of warmth in early to mid‐spring. The SI‐x products obtained using Daymet (at 1, 4, 35, and 100 km spatial resolution) and gridMET (at 4, 35, and 100 km) data, from 1980 to 2016. These products and their resulting patterns and trends across the coterminous United States, were affected more by the source of input data than by their spatial resolution. The SI‐x estimates DOY of FL (FB) are about 3 (4) weeks more accurate using Daymet than gridMET. We noted significant differences, and even contradictory rates of change in DOY, when the models were driven by Daymet or gridMET temperatures. The SI‐x products generated from gridMET poorly estimate the timing of spring onset, whereas Daymet SI‐x products and actual volunteered observations are moderately correlated (r = 0.7). Daymet better captures temperature regimes, particularly in the western United States, and hence it is more appropriate for generating high‐resolution SI‐x products at continental scale.
The main conclusion of this PhD thesis is that a careful analysis of phenological data and models is necessary to avoid misleading results. Geocomputational principles and approaches provide an ideal paradigm to integrate and scale up geographical and time contexts in such an analysis. The illustrated workflows reveal that variations in the consistency and source of phenological model inputs as well as the choice of phenological model significantly alter the estimation of long-term spring onset patterns and trends. Future studies that leverage volunteered observations and gridded (weather) datasets can adapt the proposed workflows to improve the robustness of their scientific analyses.
