mowerther / bnn_2022

Sentinel-3 OLCI and Sentinel-2 MSI Bayesian Neural Networks (BNNs) for oligo- and mesotrophic lakes

About

This repo contains the code to use the Bayesian Neural Networks for Sentinel-3 OLCI and Sentinel-2 MSI as described in:
A Bayesian approach for remote sensing of chlorophyll-a and associated retrieval uncertainty in oligotrophic and mesotrophic lakes.

BNNs were developed for oligotrophic and mesotrophic lakes (maximum chla 68 $\text{mg m}^{-3}$) and for optical water types (OWTs) [2, 3, 4, 5, 9] described in Spyrakos et al. (2018). Usage outside of this application scope is not intended. The OWT classification is implemented in this repository and an OWT flag generated for each input observation.

Figure 14 from the manuscript: OLCI BNN chla and uncertainty products over southern New Zealand, 9th of February 2020. Chla was measured in situ at the location of Lake Hawea (star symbol) as 0.89 during the overpass.

Installation

You want to have Anaconda installed and use a dedicated Anaconda environment. To install all required packages and versions, follow these steps (recommended):

Clone this repository (e.g., through GitHub Desktop). Alternatively download the repository.
Open your shell (e.g. CMD) and navigate to the directory where you cloned or downloaded the repository to, for example: cd C:\github_repos\BNN_2022

Then:
conda create -n "bnn_2022_env" python=3.8.15. This creates a fresh conda environment with the correct Python version to load the BNN models. Activate it: conda activate bnn_2022_env.
Then: pip install pandas tqdm tensorflow tensorflow_probability numpy uncertainty-toolbox xarray matplotlib. This installes all the repository requirements into your "bnn_2022_env" environment. You can then use any Python IDE, such as VSCode, activate/set the conda environment “bnn_2022_env” and navigate to the folder of this repository on your local drive.

Usage

The scripts bnn_insitu_dataframe.py and bnn_netcdf_dataset.py come with example data and necessary processing steps. I suggest to try running these successfully before using own data.

The input data from a satellite sensor should be the remote sensing reflectance $(\text{sr}^{-1})$ derived through prior atmospheric correction. The BNNs can also be applied to in situ remote sensing reflectance observations that must correspond to the relative spectral response (RSR) of OLCI or MSI S2A/S2B, and measured as or transformed to above-water reflectance. Negative values < 665 nm will be flagged, and a warning issued. The code raises ValueErrors when observations contain NaN values.

Happy usage! Any bugs reported in the issues tab or send me an email directly: mortimer(dot)werther(at)eawag.ch.