The goal of this repo is to provides some examples for performing deep learning for scene classification, semantic segmentation, and instance segmentation applied to geospatial data using fast.ai and the arcgis.learn module. Fast.ai allows for easier implementation of PyTorch while the arcgis.learn module makes use of fast.ai and makes it easier to prepare training data, train models on geospaital data, and predict back to and assess on spatial datasets. The documentation for these tools can be found here:
Note that the arcgis.learn module currently uses fast.ai v1.
Example notebooks for implementing deep learning using the arcgis.learn module can be found here. A book on fast.ai is avaialble here.
For simplicity, all examples will assume you are using a Python environment installed along with ArcGIS Pro. Deep learning tools are not available by default in the Python environment that installs with ArcGIS Pro. However, the environment can be updated to provide this functionality. This process is explained at this link and is demonstrated in the video below. This process works for ArcGIS Pro version 2.6 and greater.
The video below provides offers an introduction to deep learning as applied to geospatial data.
Lecture modules associated with these examples are avaialble on the WV View webpage as part of the remote sensing course class. Example datasets generated by us are available under Research on the West Virginia View webpage. Links to other datasets used in the examples will be provided in those specific examples.