Sen4AgriNet

A Sentinel-2 multi-year, multi-country benchmark dataset for crop classification and segmentation with deep learning

Contributors: Sykas D., Zografakis D., Sdraka M.

---

Supplementary repo with DL experiments using the Sen4AgriNet dataset: Sen4AgriNet-Models.

---

This repository provides a native PyTorch Dataset Class for Sen4AgriNet dataset (patches_dataset.py). Should work with any new version of PyTorch1.7.1+ and Python3.8.5+.

Dataset heavily relies on cocoapi for dataloading and indexing, therefore make sure you have it installed:

pip3 install pycocotools

Then make sure every other requirement is installed:

pip3 install -r requirements.txt

Instructions

In order to use the provided PyTroch Dataset class, the required netCDF files of Sen4AgriNet must be downloaded and placed inside the dataset/netcdf/ folder. These files are available for download at Dropbox, Google Drive and HuggingFace Hub (experimental).

Then, three separate COCO files must be created: one for training, one for validation and one for testing. Alternatively, the predefined COCO files for the 3 Scenarios can be downloaded from here.

After this initial setup, patches_dataset.py can be used in a PyTorch deep learning pipeline to load, prepare and return patches from the dataset according to the split dictated by the COCO files. This Dataset class has the following features:

• Reads the netCDF files of the dataset containing the Sentinel-2 observations over time and the corresponding labels.
• Isolates the Sentinel-2 bands requested by the user.
• Computes the median Sentinel-2 image on a given frequency, e.g. monthly (or loads precomputed medians, if any).
• Returns the timeseries of median images inside a predefined window.
• Normalizes the images.
• Returns hollstein masks for clouds, cirrus, shadow or snow.
• Returns a parcel mask: 1 for parcel, 0 for non-parcel.
• Can alternatively return binary labels: 1 for crops, 0 for non-crops.

Dataset exploration

This is roughly the way that our patches_dataset.py works. The whole procedure is also described in the provided notebook.

1. Open a netCDF file for exploration.

import netCDF4
from pathlib import Path

patch = netCDF4.Dataset(Path('data/2020_31TCG_patch_14_14.nc'), 'r')
patch

Outputs

"""
<class 'netCDF4._netCDF4.Dataset'>
root group (NETCDF4 data model, file format HDF5):
    title: S4A Patch Dataset
    authors: Papoutsis I., Sykas D., Zografakis D., Sdraka M.
    patch_full_name: 2020_31TCG_patch_14_14
    patch_year: 2020
    patch_name: patch_14_14
    patch_country_code: ES
    patch_tile: 31TCG
    creation_date: 27 Apr 2021
    references: Documentation available at .
    institution: National Observatory of Athens.
    version: 21.03
    _format: NETCDF4
    _nco_version: netCDF Operators version 4.9.1 (Homepage = http://nco.sf.net, Code = http://github.com/nco/nco)
    _xarray_version: 0.17.0
    dimensions(sizes):
    variables(dimensions):
    groups: B01, B02, B03, B04, B05, B06, B07, B08, B09, B10, B11, B12, B8A, labels, parcels
"""

2. Visualize a single timestamp.

import xarray as xr

band_data = xr.open_dataset(xr.backends.NetCDF4DataStore(patch['B02']))
band_data.B02.isel(time=0).plot()

3. Visualize the labels:

labels = xr.open_dataset(xr.backends.NetCDF4DataStore(patch['labels']))
labels.labels.plot()

4. Visualize the parcels:

parcels = xr.open_dataset(xr.backends.NetCDF4DataStore(patch['parcels']))
parcels.parcels.plot()

5. Plot the median of observations for each month:

import pandas as pd
# Or maybe aggregate based on a given frequency
# Refer to
# https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#timeseries-offset-aliases
group_freq = '1MS'

# Grab year from netcdf4's global attribute
year = patch.patch_year

# output intervals
date_range = pd.date_range(start=f'{year}-01-01', end=f'{int(year) + 1}-01-01', freq=group_freq)

# Aggregate based on given frequency
band_data = band_data.groupby_bins(
    'time',
    bins=date_range,
    right=True,
    include_lowest=False,
    labels=date_range[:-1]
).median(dim='time')

If you plot right now, you might notice that some months are empty:

(Optional) Fill in empty months:

import matplotlib.pyplot as plt

band_data = band_data.interpolate_na(dim='time_bins', method='linear', fill_value='extrapolate')

fig, axes = plt.subplots(nrows=3, ncols=4, figsize=(18, 12))

for i, season in enumerate(band_data.B02):

    ax = axes.flat[i]
    cax = band_data.B02.isel(time_bins=i).plot(ax=ax)


for i, ax in enumerate(axes.flat):
    ax.axes.get_xaxis().set_ticklabels([])
    ax.axes.get_yaxis().set_ticklabels([])
    ax.axes.axis('tight')
    ax.set_xlabel('')
    ax.set_ylabel('')
    ax.set_title(f'Month: {i+1}')


plt.tight_layout()
plt.show()

PatchesDataset usage example

Please refer to the provided notebook for a detailed usage example of the provided PatchesDataset.

1. Read the COCO file to be used.

from pathlib import Path
from pycocotools.coco import COCO
root_path_coco = Path('coco_files/')
coco_train = COCO(root_path_coco / 'coco_example.json')

2. Initialize the PatchesDataset.

from torch.utils.data import DataLoader
from patches_dataset import PatchesDataset
from utils.config import LINEAR_ENCODER
root_path_netcdf = Path('dataset/netcdf')  # Path to the netCDF files
dataset_train = PatchesDataset(root_path_netcdf=root_path_netcdf,
                               coco=coco_train,
                               group_freq='1MS',
                               prefix='test_patchesdataset',
                               bands=['B02', 'B03', 'B04'],
                               linear_encoder=LINEAR_ENCODER,
                               saved_medians=False,
                               window_len=6,
                               requires_norm=False,
                               return_masks=False,
                               clouds=False,
                               cirrus=False,
                               shadow=False,
                               snow=False,
                               output_size=(183, 183)
                              )

3. Initialize the Dataloader.

dataloader_train = DataLoader(dataset_train,
                              batch_size=1,
                              shuffle=True,
                              num_workers=4,
                              pin_memory=True
                             )

4. Get a batch.

batch = next(iter(dataloader_train))

The batch variable is a dictionary containing the keys: medians, labels, idx. batch['medians'] contains a pytorch tensor of size [1, 6, 3, 183, 183] where:

• batch size: 1
• timestamps: 6
• bands: 3
• height: 183
• width: 183

batch['labels'] contains the corresponding labels of the medians, which is a pytorch tensor of size [1, 183, 183] where:

• batch size: 1
• height: 183
• width: 183

batch['idx'] contains the index of the returned timeseries.

Webpage

Dataset Webpage: https://www.sen4agrinet.space.noa.gr/

Experiments

Please visit Sen4AgriNet-Models for a complete experimentation pipeline using the Sen4AgriNet dataset.

Citation

To cite please use:

@ARTICLE{
  9749916,
  author={Sykas, Dimitrios and Sdraka, Maria and Zografakis, Dimitrios and Papoutsis, Ioannis},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  title={A Sentinel-2 multi-year, multi-country benchmark dataset for crop classification and segmentation with deep learning},
  year={2022},
  doi={10.1109/JSTARS.2022.3164771}
}