Description

Deep learning for multi-modal classification of cloud, shadow and land cover scenes in high-resolution satellite imagery implemented using Keras as described in: Shendryk, Y., Rist, Y., Ticehurst, C. and Thorburn, P. (2019). "Deep learning for multi-modal classification of cloud, shadow and land cover scenes in PlanetScope and Sentinel-2 imagery." ISPRS Journal of Photogrammetry and Remote Sensing 157: 124-136.

Acknowledgements

This work exists thanks to:

1. Yannik Rist
2. Catherine Ticehurst
3. Peter Thorburn

and CSIRO's Digiscape Future Science Platform

Requirements

• conda
• git

Installation

1. Clone the repository.

   cd path/to/where/project/lives
   git clone --recursive https://github.com/yurithefury/ChipClassification.git

2. Conda environment install on local machine conda-env create -f env.yaml will create an appropriate environment called 'keras'. In one already exists, you will have to edit the first line of env.yaml to something else, eg name: veryniceenvironment. If your computer does NOT have a GPU, you will have to edit env.yaml to change the following line - keras-gpu=2.2.4 to read - keras=2.2.4

Project Contents

create_final_combined_dataset.py

This script creates the datasets for training the final models used for inference.

inference_example.py

This is an example of how to perform inference using a single Keras module and the utils.inference module. It works on TIF scenes. If you run this you should be able to look at the output in the scrap/ subdirectory.

nanscan.py

A small utility script to check for NaNs in a HDF5 file.

plot_training_history.py

Exactly what you'd think.

test_keras.sh

A Bash script for testing whether your environment works and whether Keras can find your GPUs.

train_sugarbyte.py

A training script, for use with DistributedDuctTape, but can also be run as a standalone script.

qgis/

Contains some QGIS style-files for visualisation of outputs.

models/

Inference-ready models live in here.

hdf5cvtools/

This is a collection of functions for turning geospatial raster data into HDF5 datasets suitable for machine learning tasks. It is a one-file module. See the source code for docs.

data/

Here you can find T-PS and T-S2 datasets, which is also available at Mendeley Data. A-PS data could be found at Planet: Understanding the Amazon from Space