Awesome

Cloud-Segmentation on satellite imagery data from the Sentinel-2 mission.

Problem Description

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.

Dataset

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".

Band	Description	Center wavelength
B02	Blue visible light	497 nm
B03	Green visible light	560 nm
B04	Red visible light	665 nm
B08	Near infrared light	835 nm

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

Getting Started

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

Results

Model Name	Public mIoU Score	Private mIoU Score
DeepLabV3Plus with ResNet101 as backbone	0.8805	0.8775
Unet with InceptionV4 as backbone	0.8776	0.8749
DeepLabV3 with ResNet101 as backbone	0.8299	0.8340

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.

People

Vidit Agarwal	Vedant Kaushik	Utkarsh Pandey

https://github.com/Viditagarwal7479	https://github.com/vedantk-b	https://github.com/Kratos-is-here