Package overview#

What is DeepForest?#

DeepForest is a python package for training and predicting ecological objects in airborne imagery. DeepForest comes with prebuilt models for immediate use and fine-tuning by annotating and training custom models on your own data. DeepForest models can also be extended to species classification based on new data. DeepForest is designed for:

  1. Applied researchers with limited machine learning experience

  2. Applications with limited data that can be supported by prebuilt models

  3. Scientists looking for an easy-to-use baseline to compare methods against

DeepForest uses deep learning object detection networks to predict the location of ecological objects in airborne imagery. The design of DeepForest is intended to be simple, modular, and reproducible.

../_images/image.png

For more about the motivation behind DeepForest, see some recent talks we have given on computer vision for ecology and practical applications to machine learning in environmental monitoring.

Airborne Ecology#

Practical Intro to Computer Vision in Ecology Research#

Where can I get help, learn from others, and report bugs?#

Given the enormous array of forest types and image acquisition environments, it is unlikely that your image will be perfectly predicted by a prebuilt model. Below are some tips and general guidelines to improve predictions.

Get suggestions on how to improve a model by using the discussion board. Please be aware that only feature requests or bug reports should be posted on the issues page. The most helpful thing you can do is leave feedback on the DeepForest issue page. No feature, issue, or positive affirmation is too small. Please do it now!

Why DeepForest?#

Observing the abundance and distribution of individual organisms is one of the foundations of ecology. Connecting broad-scale changes in organismal ecology, such as those associated with climate change, shifts in land use, and invasive species require fine-grained data on individuals at wide spatial and temporal extents.

To capture these data, ecologists are turning to airborne data collection from uncrewed aerial vehicles, piloted aircraft, and earth-facing satellites. Computer vision, a type of image-based artificial intelligence, has become a reliable tool for converting images into ecological information by detecting and classifying ecological objects within airborne imagery.

There have been many studies demonstrating that, with sufficient labeling, computer vision can yield near-human-level performance for ecological analysis. However, almost all of these studies rely on isolated computer vision models that require extensive technical expertise and human data labeling. In addition, the speed of innovation in computer vision makes it difficult for even experts to keep up with new innovations.

To address these challenges, the next phase of ecological computer vision needs to reduce the technical barriers and move towards general models that can be applied across space, time, and taxa.

DeepForest aims to be simple, customizable, and modular. DeepForest makes an effort to keep unnecessary complexity hidden from the ordinary user by developing straightforward functions like “predict_tile.” The majority of machine learning projects actually fail due to poor data and project management, not clever models. DeepForest makes an effort to generate straightforward defaults, utilize already-existing labeling tools and interfaces, and minimize the effect of learning new APIs and code.

How can I contribute?#

DeepForest is an open-source python project that depends on user contributions. Users can help by:

  • Making recommendations to the API and workflow. Please open an issue for anything that could help reduce friction and improve user experience.

  • Leading implementations of new features. Check out the ‘good first issue’ tag on the repo and get in touch with the maintainers and tell us about your skills.

  • Data contributions! The DeepForest backbone tree and bird models are not perfect. Please consider posting any annotations you make on Zenodo, or sharing them with DeepForest maintainers. Open an issue and tell us about the RGB data and annotations. For example, we are collecting tree annotations to create an open-source benchmark. Please consider sharing data to make the models stronger and benefit you and other users.

Citation#

Most usage of DeepForest should cite two papers: The first is the DeepForest paper, which describes the Python package:

Note

Weinstein, B.G., Marconi, S., Aubry‐Kientz, M., Vincent, G., Senyondo, H. and White, E.P., 2020. DeepForest: A Python package for RGB deep learning tree crown delineation. Methods in Ecology and Evolution, 11(12), pp.1743-1751. https://doi.org/10.1111/2041-210X.13472

The second is the paper describing the particular model. See Prebuilt Setup for citations for each model.

Similar tools#

There are many open-source projects for training machine learning models. We see DeepForest as a complement to many existing and excellent packages.

  • Roboflow

    The supervision, inference and related packages within Roboflow’s ecosystem are well executed and used throughout DeepForest. The inference machine underlying Roboflow requires connection to Roboflow, a computer vision software company which requires an API key, and has a range of commercial and license structures. We think of DeepForest as a small set of curated models that are targeted towards the ecological and environmental monitoring community. Finding robust models is challenging amongst the thousands of Roboflow projects. Roboflow is designed to be an all-encompassing ecosystem, whereas DeepForest is intentionally small and aimed at existing pipelines.

  • Torchgeo

    Torchgeo is a Python library written by developers at Microsoft to help automate remote sensing machine learning. Torchgeo has general structures, but the documents and general structure are focused on raster-based remote sensing, especially using earth-facing satellite data. Torchgeo has a number of useful datasets and curates pretrained models for remote sensing applications. The Torchgeo audience is generally more experienced with machine learning than the average DeepForest user.

We hope to continue to connect with both Roboflow and Torchgeo to improve interoperability among all model types and training. The future of open-source depends on collaboration, and we welcome users from all packages to submit ideas on how best to serve the community and reduce any duplication and wasted effort. There are many packages that hold useful individual models (e.g., DetectTree2) related to individual scientific publications. Our hope with DeepForest is to wrap general routines beyond individual research projects to make machine learning applications to environmental monitoring easier.

License#

MIT License

Copyright (c) 2019, Ben Weinstein

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.