Forecasting ecosystem response to climate extremes is a crucial aspect of understanding and mitigating the impacts of climate change. However, the accuracy of current vegetation models in predicting the spatiotemporal behavior of ecosystems and their dependence on environmental factors and macro-/microclimatic conditions is limited. This is primarily due to the simplistic parametric nature of these models and the lack of precise data to constrain them. To address this limitation, we propose an innovation project aimed at developing data-driven, spatially explicit ecosystem response models using remote sensing data and conditional Generative Adversarial Networks (cGANs).
cGANs are powerful deep learning tools commonly used to generate synthetic data that closely resemble real-world data. They have achieved significant success in various domains such as image and vision computing, video and language processing. For instance, cGANs have been used to generate realistic fake true color and multispectral satellite images based on known environmental conditions or projections of urban development guided by physics-constrained background conditions.
In the DeepV project, we aim to expand the application of cGANs by developing techniques to generate synthetic satellite image time series depicting ecosystem responses. These synthetic time series will be based on globally-available Sentinel-2 remote-sensing data and environmental and hydrometeorological conditions derived from reanalysis data and topography. Our goal is to explore the use, expansion, and refinement of cGANs with innovative techniques that incorporate time series information and biophysical constraints.
To accomplish this, we will train the developed cGANs using the vast archives of Sentinel-2 satellite data, including satellite imagery, environmental data, and hydrometeorological data. By establishing a climate-to-ecosystem response rule, we can leverage these trained cGANs to simulate ecosystem responses based on background environmental conditions and projected climate extremes. Consequently, these models will serve as valuable tools for assessing ecosystem sensitivity.
To validate the effectiveness of the developed methodologies, we will apply the techniques to historical Sentinel-2 time series of test sites that were not used during the model development phase. By comparing the synthetic Sentinel-2 scenes generated by the cGANs with observed Sentinel-2 scenes, we can quantitatively evaluate the accuracy and reliability of the methodologies in an independent manner.
Once the cGANs have been created, optimized, and validated, they can be utilized to simulate ecosystem responses under different environmental and hydrometeorological conditions. This capability allows for the assessment of expected ecosystem responses over prescribed landscapes by modifying these conditions accordingly. The outcomes of this project will enhance our understanding of ecosystem dynamics and provide valuable insights for ecosystem management and climate change adaptation strategies.