NimbusNet - NCSA Hackathon III Spring 2020

NimbusNet is cloud detection project powered by various state-of-the-art deep learning algorithms.

Best Results

Our best results were obtained using a Dynamic UNet model based on ResNet-18 with Self Attention and Mish activation function. We combined spectral bands 2, 26, and 32 in the form of an RGB image and used about 52,000 blocks of 64 x 64 pixels for training. The code and results of using this model are present in Dynamic_UNet folder.

Presentation URL

https://docs.google.com/presentation/d/1Z8NWa7d3i71DCaO3xTk0UDyqK2bhMog3C_FTIU87yS0/edit?usp=sharing

File/Folder Descriptions

1. count_blocks.py

   Get some basic stats about the data. E.g., total number of blocks, the ones that are manually classified as good training data and bad training data.

2. FFN/create_dataset_FFN.py

   Create dataset for pixel-wise binary classification (Feed Forward Network (FFN) / Ensemble )

3. Dynamic_UNet/create_dataset_UNet.py

   Create dataset for UNet

4. FFN/train_FFN.py

   Train dataset using FFN

5. Dynamic_UNet/train_UNet.py

   Train dataset using UNet

6. Bagging.ipynb

   Contains ensemble of decision trees based on bagging and random patching

7. NCSA Hackathon III

   Final Presentation file

8. Multi-band_UNet contains Non-dynamic UNet model base on band 2, 26 and 31, for more details check README in the folder.