Simulating Prognostic Vegetation Phenology Using ConvLSTM

Evaluating the two-way interactions between the atmosphere and land surface processes is crucial to our understanding of regional and global climate, vegetation dynamics, and watershed hydrology. Weather and climate have deterministic effects on vegetation growth and decay. At the same time, the land surface plays a key role in partitioning radiative energy received at the Earth's surface. Vegetation, through its transpiration and evaporation, modifies the atmospheric and land surface hydrological processes.

Modeling vegetation phenology for earth system model is challenging due to its interannual variations, spatial heterogeneity, and complex ecosystem dynamics. Currently dynamic vegetation models are driven by complex ecological processes of vegetation growth and decay, carbon and nutrient cycles, ecosystem disturbances, and weather and climate; however, integrating these models into operational climate prediction systems remains a challenge. As a result, present climate models prescribe LAI distributions derived from a 20-year average climatology of the Normalized Difference Vegetation Index (NDVI). This static LAI limits models’ abilities to dynamically respond to changes in vegetation in wet and dry years. This deficiency represents a significant drawback in climate prediction efforts.

To address this deficiency, we developed a machine learning model, ML-LAI, based on convolutional Long Short-Term Memory (LSTM) techniques to study regional-scale interactions between the atmosphere and Leaf Area Index (LAI). Our model utilizes long-term NCDC summary-of-the-day station observations of precipitations and surface maximum and minimum screen-height air temperatures, as well as Moderate Resolution Imaging Spectroradiometer (MODIS) LAI product MCD15A3H, a 4-day composite dataset with a 500-meter pixel size. This product algorithm selects the best pixel available from both MODIS sensors on NASA’s Terra and Aqua satellites. Our LSTM-LAI model trained with data at 4-day intervals and 1 km spatial resolution, to predict LAI over Southwestern and south-central parts of the United States. Data from 2003 to 2021 are used for training and the model's performance for the Years 2022 and 2023 are tested. Then the machine learning model predicted LAI is evaluated against MODIS LAI product. The results show that the predicted domain-averaged LAI values differ by less than 0.03 from the actual LAI values, and effectively captured both seasonal and interannual variations. Additionally, regional LAI values based on different land cover types are analyzed. The model shows robustness in predicting LAI for shrublands, croplands, evergreen needle leaf forest, grasslands, savannas and woody savannas.

In summary, the preliminary results demonstrate its potential as an interactive vegetation model to predict LAI for climate models and to study the complex interactions between the atmosphere, biosphere, and hydrosphere.