The application of deep learning techniques to remote sensing and geospatial data is a burgeoning area of AI. However, most state-of-the-art deep learning systems have their roots in computer vision with procedures that do not necessarily lend themselves to remote sensing data. On the contrary, managing geospatial data requires handling different projections, spatial resolutions and data formats. We developed Keras Spatial, a python package for pre-processing and augmenting geospatial data. Advanced data pre-processing features of this package include accessing remote data sources directly, combining different input rasters data regardless of their native projection and resolution, and decoupling training samples dimensions from the geographic extent to open the door for prediction across different scales.
Here is a link to an experimental version of the library, which is compatible with TF V2. Also, a tutorial that provides a simple example of feeding images and labels to a toy CNN model using The tf.data API. This is a beta version, and your feedback is highly appreciated!
Reference : Soliman, Aiman, and Jeffrey Terstriep. “Keras Spatial: Extending deep learning frameworks for preprocessing and on-the-fly augmentation of geospatial data.” Proceedings of the 3rd ACM SIGSPATIAL International Workshop on AI for Geographic Knowledge Discovery. 2019.