nalankaru / multilabel-unet

Multi-label Cloud Segmentation Using a Deep Network

Multi-label Cloud Segmentation Using a Deep Network

With the spirit of reproducible research, this repository contains all the codes required to produce the results in the manuscript:

S. Dev, S. Manandhar, Y. H. Lee and S. Winkler, Multi-label Cloud Segmentation Using a Deep Network, IEEE AP-S Symposium on Antennas and Propagation and USNC-URSI Radio Science Meeting, 2019.

Please cite the above paper if you intend to use whole/part of the code. This code is only for academic and research purposes.

Code organization

The codes are written in MATLAB and python.

Usage

1. Make sure that the input images and ternary ground-truth maps are stored inside the HYTA folder. More details about the dataset can be found here. The folder structure is as follows:
   • HYTA
     • 2GT: contains the binary ground-truth images
     • 3GT: contains the ternary ground-truth images
     • images: contains all input images
     • samples: contains sample images
2. python2 training.py Run this program to train the model on the dataset of sky/cloud images with ternary ground-truth maps. It saves the trained model in ./results/model_num.json and ./results/model_num.h5. For the purpose of reproducibility, our trained model is available from this link. Please download and save the files model_num.h5 and model_num.json inside trained-model folder for subsequent evaluations.
3. python2 testing.py Run this program to generate the testing results, after model is trained. For example, this script now generates the results for B1.jpg file.
4. python2 unet_perform.py Run this program to evaluate the performance of U-Net model over 10 experiments. It saves the results in a text file ./results/unet-result.txt.
5. Open MATLAB environment, and run the script devEtal.m. It computes the performance of Dev et al. 2015 approach and saves the results in a text file ./results/devetalresult.txt.
6. Open MATLAB environment, and run the script visualize_icip.m. It visualizes the result of sample image using Dev et al. 2015 approach. For example, here we check the result for B1.jpg.