Long-term citizen science data reveals environmental controls on tropical tree flowering at the regional scale

Abstract

Tropical tree reproductive phenology is sensitive to changing climate, but inter-individual and interannual variability at the regional scale is poorly understood. While large-scale and long-term datasets of environmental variables are available, reproductive phenology needs to be measured in-site, limiting the spatiotemporal scales of the data. We leveraged a unique dataset assembled by SeasonWatch, a citizen-science phenology monitoring programme in India to assess the environmental correlates of fruiting and flowering in three ubiquitous and economically important tree species - jackfruit, mango and tamarind - in the south-western Indian state of Kerala. Using 165006 observations spread over 19596 individual trees and over 9 years, and combined temperature and rainfall predictors for the fortnight preceding each observation, accessed using the ERA5-LAND dataset, we modelled the environmental correlates of reproductive phenophases - flower and fruit occurrence - in these species using two statistical approaches - machine learning and generalised linear mixed models. In these complementary approaches, we used a leave-one-year-out approach to cross-validate the model against observations. Models of phenophase presence for all species had higher predictive power than models of phenophase intensity. We found strong influences of multiple temperature and rainfall variables - soil moisture and maximum temperature had high importance values under machine learning models while the number of consecutive dry days also had high effect sizes in the generalised linear mixed models. The effect of time-varying environmental factors like total precipitation and consecutive dry days were also modified by the static predictors like elevation, aspect and urbanisation. Taken together, our results show the pervasive influence of climate on tropical tree reproductive phenology and extent of variability among years and individuals. We also demonstrate the potential and limitations of citizen-science observations in making and testing predictions at scale for predictive climate science in tropical landscapes.