Home

Awesome

Urban Tree Detection

This repository provides code for training and evaluating a convolutional neural network (CNN) to detect tree in urban environments with aerial imagery. The CNN takes multispectral imagery as input and outputs a confidence map indicating the locations of trees. The individual tree locations are found by local peak finding. In our study site in Southern California, we determined that, using our trained model, 73.6% of the detected trees matched to actual trees, and 73.3% of the trees in the study area were detected.

Installation

The model is implemented with Tensorflow 2.4.1. We have provided an environment.yml file so that you can easily create a conda environment with the dependencies installed:

conda env create 
conda activate urban-tree-detection

Dataset

The data used in our paper can be found in a separate Github repository.

To prepare a dataset for training and testing, run the prepare.py script. You can specify the bands in the input raster using the --bands flag (currently RGB and RGBN are supported.)

python3 -m scripts.prepare <path to dataset> <path to hdf5 file> --bands <RGB or RGBN>

Training

To train the model, run the train.py script.

python3 -m scripts.train <path to hdf5 file> <path to log directory>

Hyperparameter tuning

The model outputs a confidence map, and we use local peak finding to isolate individual trees. We use the Optuna package to determine the optimal parameters of the peaking finding algorithm. We search for the best of hyperparameters to maximize F-score on the validation set.

python3 -m scripts.tune <path to hdf5 file> <path to log directory>

Evaluation on test set

Once hyperparameter tuning finishes, use the test.py script to compute evaluation metrics on the test set.

python3 -m scripts.test <path to hdf5 file> <path to log directory> 

Inference on a large raster

To detect trees in rasters and produce GeoJSONs containing the geo-referenced trees, use the inference.py script. The script can process a single raster or a directory of rasters.

python3 -m scripts.inference <input tiff or directory> \
                             <output json or directory> \
                             <path to log directory> \
                             --bands <RGB or RGBN>

Pre-trained weights

The following pre-trained models are available:

ImageryYearsBandsRegionLog Directory Archive
60cm NAIP2016-2020RGBNNorthern & Southern CaliforniaOneDrive
60cm NAIP2016-2020RGBNorthern & Southern CaliforniaOneDrive
60cm NAIP2020RGBNSouthern CaliforniaOneDrive

We also provide an example NAIP 2020 tile from Los Angeles and an example GeoJSON predictions file from the RGBN 2016-2020 model.

You can explore a map of predictions for the entire urban reserve of California (based on NAIP 2020 imagery) created using this pre-trained model.

Using your own data

To train on your own data, you will need to organize the data into the format expected by prepare.py.

Citation

If you use or build upon this repository, please cite our paper:

J. Ventura, C. Pawlak, M. Honsberger, C. Gonsalves, J. Rice, N.L.R. Love, S. Han, V. Nguyen, K. Sugano, J. Doremus, G.A. Fricker, J. Yost, and M. Ritter (2024). Individual Tree Detection in Large-Scale Urban Environments using High-Resolution Multispectral Imagery. International Journal of Applied Earth Observation and Geoinformation, 130, 103848.

Acknowledgments

This project was funded by CAL FIRE (award number: 8GB18415) the US Forest Service (award number: 21-CS-11052021-201), and an incubation grant from the Data Science Strategic Research Initiative at California Polytechnic State University.