Random forest variable importance

Two random forest algorithms to train and test feature importance of temporal, environmental, and soil physiochemical properties on CO₂flux from mangrove soils in the Red Sea under 12 hours of incubation in light conditions and, separately, under 12 hours of incubation in dark conditions. Soil properties were measured with separate soil cores collected immediately adjacent to the incubated cores. 4 replicate cores were collected each month and CO₂ flux measured by G2201-i CRDS analyser (Picarro, Santa Clara, California USA). The random forest models were built in Python v.3.9.13 and Jupyter Notebook v.6.4.12 using the RandomForestRegressor from the SciKit-Learn package v.1.0.2. Using the datasets uploaded, 80% of data was randomly selected and used for training, with the remaining 20% used for validation. To improve model performance, hyperparameter tuning was used to maximize the R² metric and 5-fold cross-validation used to assess how the model generalizes to unseen data and reduces the risk of overfitting. Subsequently, a baseline accuracy threshold was defined for feature selection, where R² ≥ 0.6 and the average 5-fold cross-validation (CV) score ≥ 0.4. Backward elimination of variables based on these performance metrics was systematically performed to maximise the number of variables included within each model without sacrificing model performance to ensure the maximum predictive power running ~120,000 model iterations. These models were used to map feature importance of the variables retained from the feature selection stage.These models were used to present the results in: J Breavington, A Steckbauer, C Fu, M Ennasri, and CM. Duarte. 'Dynamics of CO2 and CH4 fluxes in Red Sea mangrove soils'.