Meehai/neo_data_1month

NEO dataset

Preprocessing and loading scripts for NEO datasets: 1 week (TODO) and 1 month variants. Go in either 1 week or 1 month directories.

Steps:

1. Download raw data

Use the ./download_raw_data.py -o raw_data script.

For weekly data only. Use the preprocess_raw_1week_data.py raw_data -o raw_data_fixed script.

• NDVI is only sampled at 16 days, we make an average of two consecutive samples to create an 8 day average.

• We rename the data from YYYY-MM-DD to YYYY-MM-E (E=1,2,3,4 depending on which week bucket). This is needed because different sensors sample data at different weeks

2. Convert from PNG to NPY files

./convert_png_to_npy.py raw_data -o npy_data [--resolution 540 1080]

Converts from raw PNG to [0:1] normalied, 540x1080 npy files, and adds NaNs to invalid regions in the NEO data.

If --resolution is not set, it will default to 540 x 1080.

3. Create train and test set

Run the following script:

./scripts/symlink_train_val_split.py neo_1month/data -o neo_1month/data_split --mode train_set
./scripts/symlink_train_val_split.py neo_1month/data -o neo_1month/data_split --mode test_set
./scripts/symlink_train_val_split.py neo_1month/data -o neo_1month/data_split --mode train_set_plus_missing # optional
# copy stats files too :)
cp neo_1month/data/.task_statistics.npz neo_1month/data_split/train_set
cp neo_1month/data/.task_statistics.npz neo_1month/data_split/test_set
cp neo_1month/data/.task_statistics.npz neo_1month/data_split/train_set_plus_missing

4. Run the provided viewer

Run the neo_viewer and data analysis notebook

• Set the path to neo_1week/npy_data or neo_1month/npy_data.

5. Use in ML python scripts

Use the neo reader like in the viewer above.

from neo_reader import MultiTaskDataset, neo_task_types
reader = MultiTaskDataset(dataset_path, handle_missing_data="fill_none", task_types=neo_task_types)
train_data = train_reader[0:5][0] # a batch of first 5 items. [0] is the raw torch data as a dict
pprint(train_data)
>> {
 'AOD': tensor[5, 540, 1080, 1] n=2916000 (11Mb) x∈[0.004, 0.996] μ=0.223 σ=0.237 NaN!,
 'CHLORA': tensor[5, 540, 1080, 1] n=2916000 (11Mb) NaN!,
 'CLD_FR': tensor[5, 540, 1080, 1] n=2916000 (11Mb) NaN!,
 'CLD_RD': tensor[5, 540, 1080, 1] n=2916000 (11Mb) x∈[0.024, 0.996] μ=0.546 σ=0.192 NaN!,
 'CLD_WP': tensor[5, 540, 1080, 1] n=2916000 (11Mb) NaN!,
 'COT': tensor[5, 540, 1080, 1] n=2916000 (11Mb) x∈[0.004, 0.996] μ=0.375 σ=0.220 NaN!,
 'CO_M': tensor[5, 540, 1080, 1] n=2916000 (11Mb) x∈[0.063, 0.996] μ=0.449 σ=0.168 NaN!,
 'FIRE': tensor[5, 540, 1080, 1] n=2916000 (11Mb) x∈[0.157, 0.996] μ=0.846 σ=0.148 NaN!,
 'LAI': tensor[5, 540, 1080, 1] n=2916000 (11Mb) x∈[0.004, 0.996] μ=0.314 σ=0.298 NaN!,
 'LSTD': tensor[5, 540, 1080, 1] n=2916000 (11Mb) x∈[0.004, 0.996] μ=0.632 σ=0.275 NaN!,
 'LSTD_AN': tensor[5, 540, 1080, 1] n=2916000 (11Mb) x∈[0.004, 0.996] μ=0.545 σ=0.175 NaN!,
 'LSTN': tensor[5, 540, 1080, 1] n=2916000 (11Mb) x∈[0.004, 0.996] μ=0.493 σ=0.248 NaN!,
 'LSTN_AN': tensor[5, 540, 1080, 1] n=2916000 (11Mb) x∈[0.004, 0.996] μ=0.531 σ=0.167 NaN!,
 'NDVI': tensor[5, 540, 1080, 1] n=2916000 (11Mb) x∈[0.004, 0.996] μ=0.481 σ=0.297 NaN!,
 'NO2': tensor[5, 540, 1080, 1] n=2916000 (11Mb) NaN!,
 'OZONE': tensor[5, 540, 1080, 1] n=2916000 (11Mb) NaN!,
 'SNOWC': tensor[5, 540, 1080, 1] n=2916000 (11Mb) x∈[0.004, 0.996] μ=0.775 σ=0.336 NaN!,
 'SST': tensor[5, 540, 1080, 1] n=2916000 (11Mb) NaN!,
 'WV': tensor[5, 540, 1080, 1] n=2916000 (11Mb) x∈[0.004, 0.996] μ=0.408 σ=0.310 NaN!
}