Skip to content

Treepedia 2.0: Deep Learning Based Large Scale Quantification of Urban Canopy Cover

Notifications You must be signed in to change notification settings

diegomanya/treepedia_dl_public

 
 

Repository files navigation

Fast and Accurate Estimation of Green View Index

Replication of results for the following paper: Treepedia 2.0: Deep Learning Based Large Scale Quantification of Urban Canopy Cover, which is currently published as an arxiv preprint, and has been accepted/will appear in the IEEE BigData Congress 2018 Conference Proceedings.

This project is by the Treepedia team at the MIT Senseable City Lab. It is also meant to be an extension of the existing Treepedia repository by directly replacing the GreenView_Calculate.py script. Further details are provided below.

We would love to hear how this has helped in your own projects (contact us at [email protected]), and would appreciate (but not require) citation and attribution if you have used our code/models in any of your projects.

Requirements and setup

We recommend using Python 3.5. There have been documented problems of users who load weights in Python 3.6 from models trained in Python 3.5 (see issue).

Packages that are used include keras, tensorflow (or tensorflow_gpu), PIL and scikit-image. You can install them using pip:

pip install tensorflow keras Pillow scikit-image

After doing so, you can clone this repository and also download the trained model weights and Google Street View (GSV) images. Copyright belongs to Google for original images, ground truth images are licensed under the MIT license, copyright belongs to the original authors of the paper - Bill Cai, Xiaojiang Li, Ian Seiferling and Carlo Ratti.

git clone [email protected]:billcai/treepedia_dl_public.git

Google Drive link for trained model weights: Download here

Google Drive link for GSV images: Download here

You may use other python packages (Google API Client, pydrive, gdrive) to download the files from the links above.

Next, unzip the files into the cloned treepedia-github repository. Make sure that the DOWNLOADED-FILES file is in the root folder of the cloned repo. You are now ready!

Replicating results

Simply run

python rep_results.py

This script runs the direct end-to-end GVI model, where a ResNet (residual network) was trained to directly estimate the Green View Index. A sample output of running the above command is:

Average Absolute GVI error is: 0.047427
5-95 percentile of Absolute GVI error is (-0.098191, 0.083860)
Correlation of predicted and true GVI is: 0.939691

Replacing GreenView_Calculate.py

We replaced the VegetationClassification function in the old GreenView_Calculate.py with predict_single function. Results in the paper show that we should expect increase in evaluation speed and Green View Index (GVI) accuracy.

If you wish to incorporate predict_single into your own script, predict_single has the following arguments:

def predict_single(model, image, verbosity)

model is a loaded Keras model, which can be loaded from a hdf5 saved file using keras.models.load_model. image is a numpy array that must be in the format of [H,W,3]. verbosity determines what is printed; verbosity=1 means that evaluation time estimates would be printed as well.

About

Treepedia 2.0: Deep Learning Based Large Scale Quantification of Urban Canopy Cover

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%