Skip to content
No description, website, or topics provided.
Python Shell
Branch: master
Clone or download
Type Name Latest commit message Commit time
Failed to load latest commit information.
data adding elevation file Oct 2, 2017
srcnn-tensorflow @ 29388fe initializing code, read/write needs work Sep 29, 2017
.gitignore initializing code, read/write needs work Sep 29, 2017
.gitmodules initializing code, read/write needs work Sep 29, 2017 quick look Oct 2, 2017
config-large.ini config file for optimal downscaling Oct 2, 2017
deepsd-network.png Adding figures of model Oct 2, 2017
deepsdflow.png Adding figures of model Oct 2, 2017 config file includes all parameters Oct 2, 2017
res.pdf running example Oct 2, 2017 experiment number -> model number Oct 2, 2017 end to end example runs Sep 30, 2017 end to end example runs Sep 30, 2017 experiment number -> model number Oct 2, 2017

DeepSD: Generating High Resolution Climate Change Projections through Single Image Super-Resolution

The code in this repository contains the Stacked Super-resolution CNN proposed in our recent KDD paper, DeepSD.

When cloning the repository, make sure to use the --recursive argument as DeepSD directly relys on our SRCNN library. Ie. git clone --recursive

SRCNN with AUX Var Network Architecture

Alt text

DeepSD is a Stack of SRCNNs w/ AUX var

Alt text


The current codebase has only been tested with Python2, not Python3. Major dependencies include Tensorflow and Gdal libraries. The user must install these independently of this package. I will try to add more support regarding dependencies at a later time.


Quick Look

To run the example code execute the following on the terminal:

Configuration File

config.ini provides and example configuration file allowing one to selection prism data options, file directories, architecture selection, and how many models to stack. All of the following scripts which download and process the data, train the models, and do inference, rely directly on this configuration file. One can make their own config file and include it as an argument to the scripts.

Download and Process Data -- To simpilify the example, I download year 2014 for training and 2015 for testing (as set in config.ini), but given the high resolution, the data size is still a couple gbs. Training data is processed by selecting sub-images of size 38x38 and saved to a tfrecords file. The test set contains full prism precipitation maps and saved to corresponding tfrecord files.

Train Model -- This file handles the heavy lifting for reading in the data, building the graph, and learning the parameters. The --model_number parameter allows one to select which configured model to train, ie. what resolutions and data. Tensorboard summary files will be saved inside the configured scratch directory. The checkpoints are saved in the scratch directory which contain all then necessary information for inference. -- Edit this file to use a single script to train multiple models. Our example provides trains just two SRCNN networks.

Inference -- To downscale we need to join all the trained models to a single graph. This script loops through all the models in to freeze, join, and apply inference. This script can easily be editted to include GCM outputs.

You can’t perform that action at this time.