Welcome to bfgn
Hello there! You've found bfgn, a package for using convolutional neural networks (CNNs) in remote sensing research. If you're using remotely sensed imagery (or even large lab images) and want to use a context-based model (where multiple pixels used to make decisions), and you're looking for a way to test, compare, and deploy architectures, you're in the right place.
bfgn is an extension of ecoCNN, which was published with the paper "Uncovering ecological patterns with convolutional neural networks".
What it is
bfgn was designed as an internal research tool, one that we've now opened up in
the event that others want to use it. It's a package that manages several
common components of the cnn development pipeline in the geospatial research
context, including converting raw raster/shape files into training data,
training models (with a series of customizable templates of common network types),
the application of those models over large areas, and the generation of error reports
to assess and compare models.
What it isn't
This is not a perfected package to solve all of your geospatial CNN needs.
It is a by-product of our existing research commitments and is not guaranteed to
be bug-free, nor have extensive test coverage or exhaustive best practices.
We're working to continue and improve things, and welcome any help. See something
you think should be improved? Contribute!
Documentation
Our documentation is a work in progress, but has a solid start.
Check out the link below for a ReadTheDocs style documentation page.
Getting Started
All code is available on the main
repository, along with several
examples (more examples and tutorials to come soon).
The examples are a good way to get a handle on the workflow of bfgn.
To get going with the package, clone and install the repository -
we recommend using the provided conda environment
files (available for both gpu and cpu environments). Installation details are on
the repository home page.
We're academics, and citations help us keep this work going. If you use this code
in your academic work, please cite us! When the package develops a bit further,
we'll add version-specific citations, but for now, please use:
Philip G. Brodrick, Andrew B. Davies, and Gregory P. Asner. "Uncovering
Ecological Patterns with Convolutional Neural Networks." Trends in ecology
& evolution (2019).
About Us
This code base was initially developed by a joint effort between
Phil Brodrick and
Nick Fabina, working out of the
Center for Global Discovery and Conservation Science
at ASU.