NAIP - CNN

Predicting moderate resolution forest attributes (e.g. cover and height) from high-resolution NAIP aerial imagery using ConvNets.

Usage

See CONTRIBUTING.md for Docker setup instructions.

Generating Training Data

Training data is sampled from NAIP imagery and LiDAR acquisitions stored on Google Earth Engine. Run notebooks/01_generate_training_data.ipynb to export and process tabular training data.

Model Training

Edit src/naip_cnn/cli/config.py to specify the training data and model parameters. Run the following command to train a model:

naip_cnn train 

The trained model, configuration, and summary metrics will be logged to Weights & Biases.

Inference

To predict raster attributes from a trained model stored on Weights & Biases, first download the corresponding NAIP as a TFRecord dataset (and an optional GeoTIFF mask) by running notebooks/02_export_naip.ipynb. Then run:

naip_cnn predict [MODEL_PATH] [DATASET_ID] [DATASET_YEAR]

For example:

naip_cnn predict aazuspan-team/naip-cnn/dmrpcb14 MALHEUR 2009

The output GeoTIFF will be written to data/pred/.

If you exported a mask alongside the TFRecord, use the --apply-mask option to set a NoData mask in the prediction.