Decoding tropical deciduousness phenology : patterns, drivers and implications to regional carbon cycle

Abstract

Tropical rainforests, known for being evergreen, importantly regulate large-scale carbon and water fluxes and regional climate. Despite landscape-level evergreen, increasing fine-scale studies reveal massive “non-green” (deciduousness) phenological dynamics within tropical ecosystems, resulting from the coexistence of evergreen and deciduous trees. This cryptic phenology pattern of “landscape evergreen yet strong fine-scale seasonality” highlights an urgent need for accurate characterizations of tropical deciduousness phenology across scales, from individual trees to ecosystems, which remains challenging. To resolve the knowledge gaps, this thesis integrates deep learning methods with multisource remote sensing observations to enable tropical deciduousness monitoring from individual to ecosystem scales. It also evaluates the impact of spatial scale on tropical deciduousness monitoring using multi-source satellite observations and determines the critical scale for extensive monitoring in tropical regions. Finally, the thesis examines the geographical variation of tropical deciduous phenology at regional scale and investigates its drivers and implications for the large-scale carbon cycle. The research first integrates deep learning methods with phenocam image time series to classify leaves and non-leaves, deriving leaf fraction (the opposite of deciduousness) at the crown scale. The results demonstrate that the deep learning model can accurately capture leaf fraction seasonality at the crown scale with phenocam observations. Integrating leaf fraction and greenness metrics can improve the characterization of tropical leaf phenology dynamics at fine scales, such as inferring changes in leaf age. Subsequently, the thesis integrates the spectral unmixing method with PlanetScope satellites with 3m resolution to upscale tropical deciduousness monitoring from limited sites to extensive landscapes. An ecological knowledge guided deep learning model (IG-ECAE) is developed here, which allows for automatic extraction of pure endmember spectra (critical to spectral unmixing models) and enables automatic monitoring of tropical deciduousness phenology across various landscapes. The results show that the IG-ECAE model can accurately track spatial and temporal variations of deciduousness at 16 representative tropical forest sites. To identify cost-effective methods for extensive monitoring across tropics, this research further evaluates the impacts of spatial resolution on tropical deciduousness monitoring using multi-source satellite observations and the IG-ECAE model. The results reveal that finer spatial resolutions yield higher accuracies for deciduousness monitoring. Further, Sentinel-2 data provides satisfactory monitoring of ecosystem-scale deciduousness seasonality, although less accurate than PlanetScope in capturing fine-scale phenological diversity. Lastly, the thesis extends the developed method to the regional scale for characterizing the deciduousness phenological patterns across the Amazon forests, revealing significant spatial variability in deciduousness amplitude, with 69% of the area exhibiting an amplitude more than 10%. Integrating both “green” and “non-green” metrics can improve the interpretation of tropical phenology dynamics and enable more accurate simulations of carbon fluxes. Ultimately, the results show that deciduousness variability is jointly regulated by hydroclimate, soil fertility, and tree characteristics, reflecting the resource acquisition and drought avoidance strategies in regulating such variability. Collectively, this thesis provides a bottom-up approach to quantify the deciduousness phenology across tropical regions and provides a new paradigm in tropical forest research to incorporate non-green phenology, enabling more accurate assessments of tropical forests’ response to climate change.