Awesome

crop-type-mapping

Crop type mapping of small holder farms in Ghana and South Sudan

INSTALLATION INSTRUCTIONS

Install Python 3.6

Install conda and build the environment with the following command:

conda env create -f environment.yaml

DATASET / ENVIRONMENT SETUP

These datasets are now available for free on Radiant Earth's MLHub, and through Sustain Bench.

Radiant Earth MLHub

• The dataset for Ghana is here: http://registry.mlhub.earth/10.34911/rdnt.ry138p/
• The dataset for South Sudan is here: http://registry.mlhub.earth/10.34911/rdnt.v6kx6n/
• The dataset files are saved as tifs, and will need to be restructured to work as input to the model, which initially used an hdf5 file.

Sustain Bench

• Both dataset are available via pytorch data loaders
• See more information here: https://sustainlab-group.github.io/sustainbench/docs/datasets/sdg2/crop_type_mapping_ghana-ss.html

RUN INSTRUCTIONS

To visualize training, open a separate terminal and run the following before running the main training code:

python -m visdom.server

Replace “localhost” with the static IP address provided on google cloud

To start training models, use the train.py script in the root directory of the code.

Example for CLSTM-only network: python train.py --model_name=only_clstm_mi --country=southsudan --var_length --name=southsudan_clstmonly --env_name=myenv --dataset=full --epochs=130 --batch_size=5 --optimizer=adam --lr=0.003 --weight_decay=0 --loss_weight=True --weight_scale=1 --seed=1 --s2_num_bands=10 --dropout=0.5 --clip_val=True

Example for 3D UNet model: python train.py --model_name=unet3d --country=southsudan --num_timesteps=24 --lr=0.0003 --s2_agg=False --include_indices=True --include_doy=True --use_planet=True --planet_agg=False --name=southsudan_3dunet_use_planet_noagg --env_name=myenv --dataset=full --epochs=130 --batch_size=5 --optimizer=adam --weight_decay=0 --loss_weight=True --weight_scale=1 --seed=1 --s2_num_bands=10 --dropout=0.5 --clip_val=True --hidden_dims=128

Example for Multi-Input 2D UNet + CLSTM model: python train.py --model_name=mi_clstm --country=ghana --var_length --main_crnn=True --early_feats True --include_indices=True --include_doy False --sample_w_clouds False --include_clouds False --lst_cloudy False --use_planet=True --resize_planet=False --name=ghana_use_planet_highres --env_name=ghana_use_planet_highres --use_s1=True --dataset=full --epochs=130 --batch_size=5 --optimizer=adam --lr=0.003 --weight_decay=0 --loss_weight=True --weight_scale=1 --seed=1 --s2_num_bands=10 --dropout=0.5 --clip_val=True

Example for Multi-Input 2D UNet + CLSTM earlier fused model: python train.py --model_name=fcn_crnn --country=southsudan --name=southsudan_fcn_crnn --env_name=myenv --dataset=full --epochs=130 --batch_size=5 --optimizer=adam --lr=0.001 --weight_decay=0 --loss_weight=True --weight_scale=1 --seed=1 --s2_num_bands=10 --dropout=0.5 --clip_val=True --include_s1=True

Model type and country are set with the model_name and country flags, respectively. Flags also control several properties of inputs such as satellite types, temporal aggregation, Planet imagery resolution, cloud sampling, etc. Additional hyperparameter tuning settings can be set by invoking the appropriate flags. See get_train_parser in util.py for full details.