EOTorchLoader: aerial and satellite imagery datamodule for pytorch

EOTorchLoader is an open source project and Python package to simplify the use of Earth Observation imagery in deep learning and in particular for pytorch and pytorch lighnting code.

!! EARLY WORK In PROGRESS !!

Why EOTorchLoader?

Earth Observation (EO) imageries have often a large size, their height and width could easily be larger than 10k pixels, possesses more spectral bands than RVB, and also come with multi-modality : spectral imagery, SAR, height data, etc. Due to this caracteristics, EO imagery must be pre-process in order to be used in common deep learning framework as pytorch and pytorch-lightning.

One of the most common pre-processing is the chipping|tiling|croping of large EO images into small images|patchs in order to format data in the same way than natural imageries are formatted in standard deep learning datasets (ImageNet or Cityscape). It could be made offline when writing the tiling results in disk or online during the training of DL models. The latter has the advantage of not duplicated data and to enable random cropping in large imagery during training (vs fixed cropping).

Another distinctive features of EO imagery is that patchs/tiles of a large EO image could not always be seen as independants as it exists a strong spatial correlation in EOData. So in order to train a DL models on EO data and to measure is genericity this spatial correlation should be taken in account and the split between fold or train|validation|test should be made with distinct geographical area and not by random splitting of tile|patch.

Finally due to their multi-modal and multi-spectral aspects, EOData need specifics transforms for data augmentation in DL training, and common data augmentation libray as albumentation or kornia are useful but not sufficients for EO imagery

So EOTorchLoader aims is to bring the following features for deep learning processing of Earth observation data :

• efficient online tiling|croping of EO data in Dataloader
• enable the use of geographic split between train/val/test or kfold dataset following good pratices.
• add useful transform for EO imagery.
• configurable with hydra/omegaconf.

Documentation

Learn more about EOTorchLoader in its official documentation at https://ndavid.github.io/EOTorchLoader/

License

Copyright 2022 Nicolas David

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

https://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.