mirzai 0.1.0

Prediction of Exchangeable Potassium in Soil through Mid-Infrared and Deep Learning paper code

mirzai

Making the research: "Prediction of Exchangeable Potassium in Soil through Mid-Infrared Spectroscopy and Deep Learning: from Prediction to Explainability, Albinet et al., 2022" reproducible.

Making the following research paper reproducible:

"Prediction of Exchangeable Potassium in Soil through Mid-Infrared Spectroscopy and Deep Learning:from Prediction to Explainability, Albinet et al., 2022"

Link to paper (upon acceptance)

Paper with code

1. Exploratory Data Analysis (Fig. 1)

2. Data selection and transformation

3. Baseline model (PLSR):

   • Learning curve
   • Validation curve
   • Training & evaluation

4. Convolutional Neural Network (CNN):

   • Learning rate finder
   • Learning curve
   • Validation curve
   • Training & evaluation
   • Overfitting in action (Fig. 6)

5. PLSR vs. CNN figures:

   • Learning curves (Fig. 3)
   • Validation curves (Fig. 4)
   • Observed vs. predicted scatterplots (Fig. 5)
   • Global vs. local modelling (Fig. 7)

6. Interpretability

   • GradientShap values (Fig. 8) - WIP)
   • GradientShap values correlation (Fig. 9)

Setup

Getting the data

A zipped archive of the data used in this study are available for download at the following link: https://drive.google.com/file/d/1ozHZ8KHZeuaiv8lTycxe2-yo27BhFnUt/view?usp=sharing

In a local environment

The preferred way it to use Mamba. Mamba is a fast, robust, and cross-platform package manager.

To install the required dependency and proper Python version:

• Clone git clone [e-mail]:franckalbinet/mirzai.git or download the https://github.com/franckalbinet/mirzai into your local environement
• In mirzai/ root folder, execute the following Mamba command mamba env create -f environment.yml

Here below the content of mirzai/environment.yml file listing required Python version and packages:

name: mirzai
channels:
  - conda-forge
  - fastchan
  - pytorch
dependencies:
  - python=3.8
  - nbdev
  - jupyterlab
  - numpy
  - scipy
  - matplotlib=3.5.1
  - scikit-learn
  - pytorch
  - torchvision=0.12.0
  - tqdm
  - captum

• Then activate the Python environement generated: mamba activate mirzai

• And finally launch jupyter notebook

In Google Colab

...

Acknowledgements

*This work was carried out in the context of the IAEA funded Coordinated Research Project (CRP D1.50.19) titled “Remediation of Radioactive Contaminated Agricultural Land”, under IAEA Technical Contract n°23685.

We also thank Richard Ferguson from Kellogg Soil Survey Laboratory for providing access to the USDA MIR soil spectra library and the r equired training sessions for its operation.*