SEnSeIv2

Sensor Independent Cloud and Shadow Masking with Partial Labels and Multimodal Inputs - Paper here

Installation

Before installing SEnSeIv2, make sure you have a working installation of PyTorch and torchvision (with CUDA drivers, if you want to use a GPU). More info on installing pytorch here. This is not included in the requirements.txt file.

To install SEnSeIv2, with functionality as both a python package and a command line tool:

git clone git@github.com:aliFrancis/SEnSeIv2.git
cd ./SEnSeIv2
python setup.py install

This will not download the model weights from the SEnSeIv2 HuggingFace model repo. However, when first used, the weights should be downloaded automatically.

Basic Usage

Command line interface

As an example, you can produce a cloud mask (with classes clear, thin, thick and cloud shadow) for a Sentinel-2 scene

senseiv2 -vc sentinel2 <path/to/S2-scene.SAFE> <path/to/output.tif>

Or, for Landsat 8 or 9:

senseiv2 -vc landsat89 <path/to/landsat89-scene> <path/to/output.tif>

To see all options for the command line tool, which, for instance, allow you to control parameters such as the class structure of the mask, or its resolution, you can use:

senseiv2 --help

In Python

See this notebook for a more complete overview of how to use the cloud mask in python.

You can use the cloud masks within python, if you are doing your own data preprocessing, or want to customise things in other ways. A typical use-case might begin with:

from senseiv2.inference import CloudMask
from senseiv2.utils import get_model_files

scene = ... # Some numpy array representing a satellite image
descriptors = [
  {...},    # See senseiv2/constants.py for examples
  {...}
]

# Pick pre-trained model from https://huggingface.co/aliFrancis/SEnSeIv2
model_name = 'SEnSeIv2-SegFormerB2-alldata-ambiguous'
config, weights = get_model_files(model_name)

# Lots of options in the kwargs for different settings
cm = CloudMask(config, weights,verbose=True)

mask = cm(scene,descriptors=descriptors)

Advanced uses (model training etc.)

It is not easy to replicate precisely the training strategy used here, because it is not possible to redistribute all the datasets used. However, the train.py script is included, along with some sample data in samples/ to get you started. All the data used has been preprocessed into a shared format using the eo4ai tool.

After collecting some of the datasets, you can use the training script with a config file (with some modifications) from the HuggingFace repo:

python train.py path/to/config.yaml