segment eo imagery using ai

Currently implemented fastai process using fastai's unet_learner(). Similar segmentation using other deep learning libraries will be added over time.

Makes use of rschip for image preprocessing and creating normaliser.

Fully reproducible simple example including 4 band (B,G,R,NiR) Sentinel 2 image and mask image to classify water bodies.

Setup

1. Clone this repository.

2. cd to the repo and build the docker image:

docker build . --file .devcontainer/Dockerfile -t segment

3. Run the docker image ensuring access to gpu:

docker run --rm --gpus all -i -t -p 127.0.0.1:8888:8888 -w /app \
--mount type=bind,src="$(pwd)",target=/app segment

Prepare imagery

python src/preprocess/tile.py

To generate image chips for use with unet_learner.

Train a model

python src/fast_ai/train.py --image-path inputs/chips_img --epochs 2

Many more training options with defaults. Run python src/fast_ai/train.py -h to see them.

Make predictions for full image extent

(update the train-model argument to the model trained in previous step):

python src/fast_ai/predict.py \
--input-image inputs/s2_flow_country_2023_06_16_example.tif \
--trained-model models/fastai_unet_31_10_2024_1209 \
--normaliser-scaler inputs/chips_img/s2_flow_country_2023_06_16_example_normaliser.pkl \
--boundary-remove