An effective machine learning approach for predicting ecosystem CO2 assimilation across space and time

Piersilvio De Bartolomeis, Alexandru Meterez, Zixin Shu, Fanny Yang and Benjamin D. Stocker

Abstract

Accurate predictions of environmental controls on ecosystem photosyn thesis are essential for understanding impacts of climate change and extreme events on the carbon cycle and the provisioning of ecosystem services. Using time-series measurements of ecosystem fluxes paired with measurements of meteorological variables from a network of globally distributed sites and remotely sensed vegetation indices, we train a recurrent deep neural network (Long-Short-Term Memory, LSTM), a simple deep neural network (DNN), and a mechanistic, theory-based photosynthesis model with the aim to predict ecosystem gross primary production (GPP). We test the models’ ability to spatially and temporally generalise across a wide range of environmental conditions. Both neural network models outperform the theory-based model considering a leave-site-out cross-validation (LSOCV). The LSTM performs best and achieved a mean R2 of 0.78 across sites in the LSOCV and an average R2 of 0.82 across relatively moist temperate and boreal sites. This suggests that recurrent deep neural networks provide a basis for a robust data-driven ecosystem photosynthesis modelling in respective biomes. However, limits to global model upscaling are identified using a cross-validation by vegetation types and by continents and we identified weakest model performances at relatively arid sites where unknown vegetation exposure to water limitation limits model reliability.

Use

The repo contains code to run experiments on 2 models:

• FCN: A fully connected model
• LSTM: A recurrent model using and LSTM neural network

Input files

We have 2 input files:

• data/df_20210510.csv: contains the raw data, without any imputation, and is used to compute the results excluding the imputed values
• data/df_imputed.csv: contains the imputed data, used to train the model

Commands

These are the commands needed to train each of the models.

Leave-site-out

We save one CSV file, where each row contains a site and the model prediction when that site is left out of training.

LSTM Model

This model takes as input a sequence (the input data) and predicts another sequence, representing the GPP across the length of time.

python src/evaluation/leave_site_out_rnn.py

FCN Model

This model is an MLP neural network trained on the data. It was used as a baseline to compare to the rest of the methods.

python src/evaluation/leave_site_out_fcn.py

Random Forest

In the src/evaluation/random_forest directory, there is a notebook that trains a Random Forest model, also used as a baseline.

Leave-vegetation-out

Run python src/evaluation/leave_vegetation_out_rnn.py, which will output a directory named leave_{group_name}_out (where group_name corresponds to the vegetation group that the model will be trained on. In this directory, we save separate CSV files for each in-distribution site (which is left out of training) and for each out-of-distribution vegetation group.

Leave-continent-out

Run python src/evaluation/leave_continent_out_rnn.py, which will output a directory named leave_{group_name}_out (where group_name corresponds to the continent name that the model will be tested on). In this directory, we save for both the in-distribution continent and out-of-distribution continent, one CSV file per each in-distribution site (which is left out of training).

Run on cluster

To run on the cluster, execute:

bsub -R rusage[mem=10000,ngpus_excl_p=1] python SCRIPT_NAME_HERE.py

Disclaimer

For any bug reports, send an email to one of the authors.