LAI retrieval

Retrieval of leaf area index (LAI) values by inverting PROSAIL radiative transfer model. Inversion is done by training a neural network on PROSAIL simulations.

Setup the environment

pip install -r requirements.txt

1. Collect soil data across Switzerland

Using Google Earth Engine, produce yearly mean composites of bare soil pixels, for arable land. The Javascript code is provided in S2_baresoil_GEE.txt. The collected data is saved in data/S2_baresoil_GEE.zip (unzip folder before running code).
Extract a dataset of bare soils using K-means clustering and save samples to .pkl file (results/GEE_baresoil_v2/*.pkl) with

python bare_soil_GEE.py

2. PROSAIL forward runs

Launch PROSAIL forward runs with

python simulate_S2_spectra_soil.py

The runs can be configured in the file directly. (We created 50k runs with soil for Sentinel-2A and 2B each, and 50k each without soil. We then also created the test set of size 10k for Sentinel-2A and 2B.)
The parametrisation of the RTM (with and without soil) is provided in lut_params.

Results of the forward runs are stored in results/lut_based_inversion.

3. Create validation set

Download S2 data corresponding to in-situ measurements with

python download_validation.py

4. Train model

To train the inversion model you need to configure the config files (config/config_NN.yaml). You specify the model and its parameters, the train, test and validation sets.

python train.py configs/config_NN.yaml

Then test on the validation data with

python test.py configs/config_NN.yaml

5. Add noise to training data

Add noise to the LUT training data using different models and at different amounts

python noise.py

and save to results/lut_based_inversion

Train the model with different noise types and levels, and compare results saved in an excel file

python train_noise.py

6. Tune model

Perform hyperparameter tuning. Grid of hyperparmeters is provided in the config file and save performance in excel file (saved in tuning_results/)

python tune.py configs/config_NN.yaml

Then compare hyperparameter tuning results with

python compile_tuning_results.py

7. Compare to other models

Perform LUT-based inversion on the validation data, and get RMSE and $R^2$

python invert_s2_valdata.py