Abstract
Most state-of-the-art deep learning systems have their roots in computer
vision, which force the remote sensing community to develop ad-hoc
procedures for applying deep learning methods in the analysis of remote
sensing data. In this Juypuiter notebook, we present Keras Spatial
(https://pypi.org/project/keras-spatial), a new python package for
pre-processing and augmenting geospatial data for deep learning models.
Keras Spatial is composed of loosely-coupled components, which allow
users to pre-process geospatial raster data on-the-fly before ingesting
them into neural networks. The advantage of using Keras Spatial over
more traditional Ad-hoc pipelines are (1) allowing scientists and
developers to work in projected coordinates rather than pixels and (2)
controlling the sample space and hence avoiding issues such as bias and
class imbalance during training. We will demonstrate Keras Spatial using
the case study of processing digital elevation data for a segmentation
model. We will also demonstrate advanced data pre-processing features of
this package, such as accessing remote data sources directly, easy
integration of multiple datasets using automatic reprojection and
resampling, and decoupling training samples dimensions from the
geographic extent to open the door for prediction across different
scales.