Clouds can occlude objects from being detected, which hinders autonomous object detection and change detection tasks. Additionally, clouds can provide useful climate parameters. Therefore, cloud detection and segmentation is regarded as a first step in many geoscience and remote sensing applications.

Methodology and preliminary results:

The dataset used consists of Landsat images with 30m resolution and 4 channels; red, green, blue and near infrared. There are 8 scenes in total, which are cropped into multiple 384×384 images. There are 8400 images for training and 9201 images for testing. Each image has its corresponding ground truth for training and evaluation.
UNET architecture used for this task, with slight adjustments introduced to it to aid cloud segmentation. The network outputs a probability map for the whole image. Based on that, a binary mask is generated, which asserts whether a pixel is a cloud (white) or not a cloud (black). The figures below demonstrate some sample results.

Duration:

2 months (March 2020 – April 2020)

Team Members:

  • Eng. Nour Abura’ed
Fully Convolutional Neural Network (FCN) for Cloud Segmentation