To obtain adequate analytical results from multi-spectral satellite imagery, it is essential to precisely detect clouds and mask them out from any Earth surface as they obscure important ground-level features in satellite images, complicating their use in wide variety of applications from disaster management and recovery, to agriculture, to military intelligence. Thus, Improving methods of identifying clouds can unlock the potential of an unlimited range of satellite imagery use cases, enabling faster, more efficient, and more accurate image-based research.

• The challenge used publicly available satellite data from the Sentinel-2 mission, which captures wide-swath, high-resolution, multi-spectral imaging. There are four images associated with each chip. Each image within a chip captures light from a different range of wavelengths, or "band".

• For more details visit: https://www.drivendata.org/competitions/83/cloud-cover/page/398/#images

• Creating new conda environment:

    conda create -n cloudncloud -f environment.yml
    conda activate cloudncloud
    pip install segmentation_models_pytorch

• Set number of epochs, batch size, optimizer, loss function, model, transformation to be applied on data by changing them in config.py

• To use Unet with inceptionv4 as backbone

  model = smp.Unet(
               encoder_name="inceptionv4",
               in_channels=4,
               classes=2
              )

• To use DeepLabV3 with resnet101 as backbone

  model = smp.DeepLabV3(
             encoder_name="resnet101",
             in_channels=4,
             classes=2
          )

• Training and validation loops can be customised by editing

The best accuracy was achieved with DeepLabV3Plus with ResNet101 as backbone.

1st Image is a channel of the satellite image, 2nd image is true label, 3rd image is the prediction.

Vidit Agarwal	Vedant Kaushik	Utkarsh Pandey
https://github.com/Viditagarwal7479	https://github.com/vedantk-b	https://github.com/Kratos-is-here