| output | ||||
|---|---|---|---|---|
|
This repository contains all of the data and code for the project titled: "Random Forest Regression Models for Predicting Picophytoplankton Biomass Partitioning in the Eastern Pacific."
The partitioning of phytoplankton biomass is crucial for understanding how climate change impacts the carbon cycle as it sheds light on the distribution and dynamics of different phytoplankton groups. However, there is a significant knowledge gap regarding the environmental factors that drive the partitioning of phytoplankton biomass, particularly in warm, nutrient-poor regions that are predicted to expand in future ocean conditions. This study focus on the dominant phytoplankton groups in these regions: Prochlorococcus, Synechococcus, and small eukaryotic phytoplankton (less than 2 µm in diameter). Using random forest regression models, we assessed the predictability of phytoplankton biomass based on salinity, temperature, light intensity, and dissolved inorganic nutrient concentrations (nitrate, phosphorus, and iron). Model performance was evaluated using 2,800 observations per population were obtained through high-frequency flow-cytometry in surface water of the North Subtropical and Subpolar gyres. Mean Absolute Error Percentage varied among picoplankton population’s Random Forest models, Prochlorococcus: 7.71%, Synechococcus: 2.30%, Nanoeukaryotes: 3.34%, and Picoeukaryotes: 6.57%. Results highlight the importance of nitrate, salinity, and to a lesser extent, iron, as predictors of Prochlorococcus biomass. Conversely, phosphate and nitrate emerge as the primary drivers of Synechococcus and picoeukaryote biomass, respectively. The upper and lower limits of the biomass and environmental data in the dataset define the boundaries within which a Random Forest can generate its predictions, making these models specific to the location and timing of this study. Further research is needed to explore additional factors and refine prediction models by considering factors like nutrient uptake rates, grazing pressure, and microbial interactions.
All of this has been installed on MacOS systems. Ubuntu/linux and Windows users may have a slightly different experience.
Jupyter notebooks, a web-based interactive computing platform, were used as an interface for python within a conda enviorment.
A surest way to run these notebooks is use create a conda environment using the environment.yml file for this repository. On your local machine:
conda env create -f environment.yml
Once installation is complete:
conda activate machine-learning-env
jupyter notebook
If you get an error 'command not found: jupyter-notebook', first install notebook support in conda:
conda install nb_conda
Workflow organized numerically from 00 to 05 detailing every step taken on my local machine to conduct this project. Folder 04 ddetails the steps taken to split, train, and test the population specific random forest models.
A note on directories: these notebooks use git as a way to locate the nessecary files, but if you have these files outside of a repository you will need to substitute your own directory paths.
If you pulled directly from the repository, you should have all the data already and you can skip notebooks 00-03, which are querying and ingesting the data.
-
Notebook 00: run every cell as you go This whole notebook should be able to be ran without modification, but the query will take anywhere between 5-30 minutes to run depending on your internet connection and write speed. This notebook queries data from Simons CMAP Darwin climatalogical data.
-
Notebook 01: run every cell this notebook ingests PAR data and combines the SeaFlow/CMAP data and the PAR cruise data into one file.
-
Notebook 02: run every cell This notebook cleans up the combined data for use in the RF models.
-
Notebook 03: run every cell This notebook lets us take a look at the data and make sure all the data types are functional for the RF models
-
Notebook(s) 04: run each notebook in the '04_Populations-model-fitting' folder in order, starting with '01_model-preparation.ipynb'. You may have to change the file directories on 02-05 to fit your system.
-
Notebook 05: run every cell - this notebook will compare the random forest predictions to known cruise data.
Contains joblib files of the developed sklearn RandomForestRegressor() models for each plankton population. These can be pulled from this folder into your own notebook.
contains two folders, original and modified
- original: all data loaded into python notebooks that have not been previously manipulated in this project, including Simon CMAP data queried in notebook 00_cmap-querys.
- modified: all data that has been previouslt manipulated in this project, usually saved for intermidiate steps to get from one workflow to another.
- Where figures are saved throughout the notebook. All figures are displayed in the ipynb notebooks, and some exclusively so.
This work was conducted by Cristian Swift ([email protected]) under the mentorship of François Ribalet1 and François Ribalet through the School of Oceanography, Univeristy of Washington, Seattle, USA