EcoPro-LSTMv0: A Memory-based Machine Learning Approach to Predicting Ecosystem Dynamics across Time Scales in Mediterranean Environments

EcoPro-LSTM is version 0 of Ecosystem Process Modelling using a Long Short-Term Memory Approach. The model uses a Multi-Task Multi-Timescale Long Short-Term Memory (MT-LSTM) modelling framework. EcoPro-LSTM was developed by Mitra Cattry et al. as part of the research “EcoPro-LSTM v0: A Memory-based Machine Learning Approach to Predicting Ecosystem Dynamics across Time Scales in Mediterranean Environments,” accepted for publication in Machine Learning: Earth (2026). The work was financially supported by the Swiss National Science Foundation, Award #P500PN_206603, and the National Science Foundation (NSF) Science and Technology Center (STC) Learning the Earth with Artificial Intelligence and Physics (LEAP), Award #2019625-STC. The model predicts environmental variables including gross primary productivity (carbon uptake), ecosystem respiration, evapotranspiration, and surface soil moisture across multiple timescales using a single LSTM architecture, as described in the accompanying tutorial. Integrated gradients are used to enhance model interpretability by identifying the influence of climatic drivers such as precipitation. For questions, issues, or further assistance with the code, please contact Mitra Cattry at [email protected]. Citation If you use the EcoPro-LSTM model, code, processed data, or accompanying materials in your research, please cite the associated article: Cattry, M., Zhao, W., Nathaniel, J., Qiu, J., Zhang, Y., & Gentine, P. (2026). EcoPro-LSTM v0: A Memory-based Machine Learning Approach to Predicting Ecosystem Dynamics across Time Scales in Mediterranean Environments.Machine Learning: Earth. Accepted. Repository contents MT_LSTM.zip — Core model code implementing the Multi-Task Multi-Timescale LSTM framework. KG_climate.zip — Köppen–Geiger climate classifications and associated code used to identify Mediterranean and semi-arid FLUXNET2015 sites. ablation_processed_files.zip — Processed ablation results, including site-level MAE across GPP thresholds on the test sets for all model configurations, together with the Python code used to plot the results.