Skip to content

Latest commit

 

History

History
173 lines (123 loc) · 13 KB

README.md

File metadata and controls

173 lines (123 loc) · 13 KB
Raw

Poultry barn mapping

Jump to: Setup | Dataset and pretrained models | Model training and evaluation | Dataset creation and filtering

This repo contains code for training, evaluating, and running deep learning models for detecting poultry barns from high-resolution aerial imagery as well as a US-wide datasets of predicted poultry barn locations to accompany the paper "Mapping industrial poultry operations at scale with deep learning and aerial imagery". Specifically, we train and evaluate semantic segmentation models with labels from the Soroka and Duren dataset of poultry barns over the Delmarva Peninsula and aerial imagery from the National Agriculture Imagery Program (NAIP), then run our best models over large amounts of NAIP imagery to create datasets of poultry barn locations. We also develop a post-processing step to filter out false positive predictions at the barn level. Finally, we release our best trained model, two generated datasets of predicted poultry barns -- one across the entire United States with the latest 1m imagery per state, and another in the Chesapeake Bay with 2017/2018 imagery -- and a validation of our results.

If you make use of this implementation or data in your own project, or you want to refer to it in a scientific publication, please consider referencing this GitHub repository and citing our paper:

@article{robinson2021Mapping,
  title={Mapping industrial poultry operations at scale with deep learning and aerial imagery},
  author={Robinson, Caleb and Chugg, Ben and Anderson, Brandon and Ferres, Juan M Lavista and Ho, Daniel E},
  journal={arXiv preprint arXiv:2112.10988},
  year={2021}
}


Figure 1. A heatmap of predicted poultry barn locations across the United States. Download these predictions here.

Setup

First, run the following commands to create a conda environment, "cafo", with the necessary dependencies for running the scripts and notebooks in this repository:

conda env create -f environment.yml
conda activate cafo

Next, follow the instructions in the following sub-sections to set up the different datasets.

Finally, run the following notebooks to prepare various derivative files used by the training and inference pipelines:

  • notebooks/Data preparation - Generate Chesapeake Bay county GeoJSON.ipynb - This generates a GeoJSON file countaining a polygon for each county intersecting the Chesapeake Bay Watershed.
  • notebooks/Data preparation - Parse NAIP file list.ipynb - This generates a list of the most recent 1m NAIP imagery tiles per state and a list of the most recent NAIP imagery tiles per state (note: the most recent 1m imagery is usually older as most recent imagery is 0.6m).
  • notebooks/Data preparation - Prepare training and testing splits.ipynb - This generates masks for all training and testing splits.

Data sources

NAIP tiles

Use the following commands to download a list of all NAIP tiles available through the Microsoft Planetary Computer:

wget https://naipblobs.blob.core.windows.net/naip-index/naip_v002_index.zip
unzip naip_v002_index.zip
rm naip_v002_index.zip
mv naip_blob_list.txt data/

Delmarava poultry barn labels

Download the Delmarva_PL_House_Final.zip file from the Soroka and Duren 2020 Poultry barn dataset from here. From the unzipped directory run:

conda activate cafo
ogr2ogr -of GeoJSON -t_srs epsg:4326 Delmarva_PL_House_Final2_epsg4326.geojson Delmarva_PL_House_Final2.shp
ogr2ogr -of GeoJSON -t_srs epsg:32618 Delmarva_PL_House_Final2_epsg32618.geojson Delmarva_PL_House_Final2.shp
ogr2ogr -of GeoJSON -t_srs epsg:26918 Delmarva_PL_House_Final2_epsg26918.geojson Delmarva_PL_House_Final2.shp

Copy the three generated files, Delmarva_PL_House_Final2_epsg4326.geojson, Delmarva_PL_House_Final2_epsg32618.geojson, and Delmarva_PL_House_Final2_epsg26918.geojson to the data/ directory in this repository.

Dataset and pretrained models

The following are download links for our model and final generated datasets:

Each polygon in the above datasets contains 8 features:

  • p - The averaged model predicted probability over all imagery pixels within the polygon.
  • rectangle_area - The area of the entire polygon in square meters.
  • area - The area of just the positively predicted pixels under the polygon in square meters.
  • rectangle_aspect_ratio - The ratio of the polgon's long side to its short side.
  • distance_to_nearest_road - The distance from the polygon to the approximate nearest road from OpenStreetMap.
  • year - The year that the source NAIP imagery was captured.
  • date - The date that the source NAIP imagery was captured.
  • image_url - The URL to the source imagery that was used to create the prediction.

The "filtered and filtered (deduplicated) predictions" are created following the method described in Dataset creation and filtering.

Model training and evaluation

Our experiments can be reproduced with python scripts/run_experiments.py. This script will run train.py in parallel with a hyperparameter sweep. If you set TEST_MODE=True then the individual commands will be printed to STDOUT, allowing you to selectively run a subset of the experiments.

The scripts/run_test_inference_and_evaluation.py script will use the model checkpoints (the best checkpoints according to validation metrics) from each experiment, run them on all of the imagery from the test set, and report tile level metrics in CSV files. Our results from this step can be found in results/.

A notable result we find with our experiments is the importance of rotation augmentation:



Figure 2. Predictions from models trained with and without rotation augmentation. These models exhibit similar validation and test perfomance as the distribution of building rotations is similar between training/validation/test, however rotation augmentation results in a model with stronger generalization performance.

Rotation augmentation is particularly important for our models as the poultry CAFOs in our labeled data are strongly biased to have north-south oriented barns:


Figure 3. Histogram of orientations of barns from the Soroka and Duren dataset.

Dataset creation and filtering

Creation

To create the Chesapeake Bay dataset for example, first, download the model weights to output/train-all_unet_0.5_0.01_rotation_best-checkpoint.pt. Then, run the model over all NAIP imagery in the Chesapeake bay - python scripts/run_chesapeake-bay-3-18-2021_inference.py. Finally, run the post-processing pipeline over the model's per-pixel predictions - bash scripts/run_chesapeake-bay-3-18-2021_postprocessing.sh - resulting in a set of polygons with features that can be used to filter false positives as described below.

To create the dataset over the entire US we follow the same process, see scripts/.

Filtering

For filtering out false positive predictions we use the distribution of areas and aspect ratios seen in the Soroka and Duren dataset (under the assumptions that all poultry barns will follow these distributions). The mean area is 2477.18 m^2 with a standard deviation of 849.69 m^2 and a range of [525.69, 8106.53] m^2. The mean aspect ratio is 9.10 with a standard deviation of 1.72 and range of [3.4, 20.49].

Any prediction with a feature rectangle_area that falls outside of the [525.69, 8106.53] range, with a feature rectangle_aspect_ratio that falls outside of the [3.4, 20.49] range, or that has a distance_to_nearest_road of 0 is counted as a false positive and removed.

Any prediction with a feature rectangle_area that falls outside of the [525.69, 8106.53] range, with a feature rectangle_aspect_ratio that falls outside of the [3.4, 20.49] range, or that has a distance_to_nearest_road of 0 is counted as a false positive and removed.


Figure 4. Distribution of the areas and aspect ratios of barns from the Soroka and Duren dataset.

Deduplication

The NAIP tiles that we perform inference over have overlap between adjacent tiles. As such, some barns are duplicated in the raw set of predictions if they fall in these overlapping areas. To fix this, our final post-processing step is to merge overlapping predictions (i.e. to deduplicate predictions in the dataset). To do this, we iteratively merge overlapping polygons keeping the attributes of the largest. For the filtered "Full USA" set of predictions, this reduces the final number of predictions from 424,874 to 360,857 polygons.

External data licensing

License

This project is licensed under the MIT License.

The datasets are licensed under the Open Use of Data Agreement v1.0.

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.

Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.